saying that love is in our instincts not something that we grew to know

I like that he said "some standard English", I also really like the term code meshing and I agree with the concept.

Code meshing also adds individuality.

Being forced to code switch can Harm your Creativity and it ends up being sort of like a "Borrowed Voice" Instead of your own voice and creativity shining.

The idea of journalism is that it's all about these journalists expressing the truth with a moral code they have to follow. If liberalism enters the picture, then [I feel] that it'll cause a rift pitting a belief with a set of moral codes versus a belief that follows freedom and independence [since that's what I think of liberalism].

Each instance can have attributes, sometimes called instance variables. These are just like other variables in Python. We use assignment statements, with an =, to assign values to them. Thus, if alex and tess are variables bound to two instances of the class Turtle, we can assign values to an attribute, and we can look up those attributes. For example, the following code would print out 1100.

ex:

alex.price = 50 tess.price = 600 print(alex.price + tess.price)

This is a very strange, axiomatic argument. That open source should be exempt because .... open source. There are no reasons to lift regulations meant to limit harms for a certain code production model, even if it's beneficial in itself.

This reminds me of "I did write some code in Java once, but the code was in C and Lisp (I simply happened to be in Java at the time ;-)".

რას ნიშნავს receptive და productive ტიპის ლექსიკა? receptive — ფორმის აღქმა და შინაარსის გაგება. productive — სიტყვის გამოყენება წერით, ან ზეპირ მეტყველებაში.

Centralized Ethereum interface layer

Because Infura's level of abstraction is easier to code then directly on the blockchain.

• black-box penetration tests
• white-box tests
• code audits

• doGET no es async
• llama a request() sin await

this is quick checkout sample code. we should use the standard web checkout sample code

Reviewer 2. Haris Zafeiropoulos

I appreciated the opportunity to review your manuscript. Tourmaline aims at facilitating an easy-to-follow architecture for tracking input and output file names, parameters, and commands of QIIME2 runs to enhance meta-analyses. If I am not mistaken, this is the corner-stone of this study so my review is based on that. Running Tourmaline is straightforward and its documentation is exceptional. The video tutorial and the GitHub wiki (https://github.com/aomlomics/tourmaline/wiki) allows non-experienced users to start working their analysis and the containerized version of the tool allows an easy-to-go installation in multiple operating systems without extra effort. The extra visual components provide insight in a nice way and the report returned can provide added value on the runs. However, even if I do share the authors' interest on usability and interoperability and tools could have a great impact in the community indeed, Tourmaline currently lacks any substantial features to be considered as a stand-alone software tool. In addition, there are several issues that it is my belief that need to be addressed (see the following list). Major issues Major Issue #1: The authors claim that "this lack of automation and standardization [in tracking input and output file names, parameters, and commands on QIIME2] is inefficient and creates barriers to meta-analysis and sharing of results. Therefore, what Tourmaline and thus, the manuscript needs to demonstrate, is that meta-analyses are now feasible to a greater extent, thanks to the Tourmaline wrapper. Major Issue #2: Assuming that enhancing meta-analyses is the main contribution of Tourmaline, it is fundamental to consider the minimum information about a marker gene sequence (MIMARKS) standard of the Genomic Standards Consortium (GSC). Rather than just mentioning MIMARKS, Tourmaline needs to explore ways to exploit such standards, i.e. perhaps by adding MIMARKS columns in the config file. Major Issue #3: As QIIME2 has been developed on the basis of a plugin archiiated the opportunity to review your manuscript. Tourmaline aims at facilitating an easy-to-follow architecture for tracking input and output file names, parameters, and commands of QIIME2 runs to enhance meta-analyses. If I am not mistaken, this is the corner-stone of this study so my review is based on that. tecture, it would be highly recommended that such an application could be provided as a plugin too, joining the corresponding QIIME2 library (https://library.qiime2.org/plugins/). Major Issue #4: With respect to the structure of the manuscript, it is my belief that there are sections that should be omitted. Tutorials and "how to" are of extremely valuable but it would be better to be provided either as supplementary material or through repositories, e.g. GitHub wiki, GitHub pages etc, rather than in the main manuscript. The wiki page on Tourmaline's GitHub repository is rather informative. An alternative might be merging the "Overview" along with the "Snakefile", "Config file", "Input files" and "Run the workflow" sections, to describe "The Tourmaline workflow" architecture in a less verbose way, highlighting the role of the "Snakefile" and the "config.yml" files and the architecture that binds them together. "Documentation", "Installation", "Cloning" subsections could/should be omitted too. Major Issue #5: The test dataset does not allow the validation of Tourmaline in meta-analyses. It is rather important to have a testbed dataset to demonstrate "how to run" but a use case of an actual meta-analysis is required to demonstrate how different analyses can be combined in the framework of Tourmaline and provide further insight than those of the initial ones. Major Issue #7: No license has been included in the "Availability of supporting source code" section. On Tourmaline GitHub repo a license (https://github.com/aomlomics/tourmaline#license) is mentioned, yet GigaScience asks for an appropriate Open Source Initiative compliant license (https://opensource.org/licenses/category). In addition, I tried to find if the QIIME2 license is mentioned in a Tourmaline Docker container and I could not; if I am not mistaken that is required based on the QIIME2 license (https://github.com/qiime2/qiime2/blob/master/LICENSE). My apologies again in case of any misapprehension. Major Issue #8: Parameter optimization is indeed one of the greatest challenges in metabarcoding bioinformatics analyses. However, it is not clear to me how by keeping the exact same names in your output files, will you be able to compare the results of the different runs. Major Issue #9: I realise that the authors provide Figures 2 and 3 in a complementary way, presenting the visual component returned after each step. However, having a figure with 16 screenshots makes it hard for the reader to realize what is coming from QIIME2 and what from Tourmaline but most importantly does not highlight the added value that Tourmaline provides to such an analysis. It is my belief that Figure 2 could remain as it, while FIgure 3 should focus on the output components that are not provided by QIIME2 routines, but from Tourmaline functionalities. In case of a meta-analysis, this figure should highlight all the added value that using QIIME2 through the Tourmaline wrapper would provide. Minor issues Minor Issue #1: Please rephrase the Findings section in the abstract, so that it is clear that Tourmaline invokes QIIME2 routines to implement taxonomy assignment, perform analysis etc. It is required to state clearly what QIIME2 does and what are the extra features of Tourmaline throughout the manuscript. Minor Issue #2: The conclusion you mention in the abstract is not in line with the scope of Tourmaline that was described earlier. Tourmaline does not accelerate the performance of QIIME2 routines. Its aim, as mentioned earlier, is to enhance meta-analysis and sharing of results. Minor Issue #3: terms such as "meta-analysis", "reproducibility", "metadata" could be added Minor Issue #4: In line Information gained... resource management it would be nice to add references for the value of the method in each of the various fields mentioned. Minor Issue #5: Usually, (shotgun) metagenome analyses are used to measure diversity in microbiomes, meaning the functional, genomic diversity; the term microbiome has been widely used as the collection of genomes from all microbial taxa present in a sample. It would be better to rephrase this like "popular method of measuring taxonomic/microbial diversity of host microbiome or in environmental samples" Minor Issue #6: As an overall comment, long sentences make the manuscript hard to read. In this case: "PCR primers have been used to generate amplicons of the bacterial 16S rRNA gene in studies of human and animal microbiota [..] among others." should be splitted. Minor Issue #7: "other environmental surveys" please explain. Minor Issue #8: It is not clear to me how the study of Prodan et al. (2020) is related to the standardization of amplicon data analysis. Minor Issue #9: The authors highlight that the standard directory structure enhances data exploration and parameter optimization. A use case to demonstrate this main feature of Tourmaline would be of high value. Minor Issue #10: The purpose of performing amplicon sequencing or metabarcoding is to reveal patterns of diversity in biological systems. That is not the only case, please rephrase. Sincerely, Haris Zafeiropoulous

Re-review:

Most of my initial comments have been addressed to some extent. However, it is my belief that there is a major contradiction with this manuscript. As described in the introduction, Tourmaline is supposed to address challenges that make meta-analysis hard for metabarcoding studies. "; this lack of automation and standardization is inefficient and creates barriers to meta-analysis and sharing of results". However, as the authors highlight in their response, there are 5 points that make Tourmaline to set apart from other amplicon workflows. However, only one of them ("3- Snakemake features") is related (to some extent) to the challenge described. The rest are exceptional ways to make things easier for the users to run an analysis but they do not have a direct link with how to enable/support meta-analysis. Therefore, it is my belief that the Introduction section should be revised to better present the actual highlights of Tourmaline or further features (some of them described in my initial review) need to be added to support meta-analysis.

Other Issues Even though the authors recognize the impact of metadata standards: they do not mention anything on their manuscript about them and their potential I was not able to figure out how "have made the metadata that comes with Tourmaline fully MIMARKS-compliant." If this software is focused on meta-analysis, I would strongly suggest investing more effort on describing how these could benefit the community and the Tourmaline users. In the parallelization section that was added, it is fundamental to mention that this is possible thanks to Qiime2 implementation. Snakemake is working as an interface allowing Tourmaline to support the options of Qiime2. If Qiime2 had no option for running on multiple threads, then Tourmaline would not inherit such a feature. The same applies for the merging step in the meta-analysis case. All Qiime2 commands used as such need to be clear that are Qiime2 commands that are performed in the Tourmaline workflow; otherwise it can be thought that the feature was developed from the Tourmaline team

Even though I already know the answer to this question, I'm still so curious as to why it was believed that women were not capable of fighting in the "regular" army as well

Smart example of using the walrus operator :=

• source : https://www.webnots.com/how-to-share-a-webpage-in-google-chrome/

• dweb : https://bafkreih7o5jambsc2fg3uwsfd32ghojqrcssfbvnuajumg5qcr2xoefvve.ipfs.dweb.link/?filename=How%2520to%2520Share%2520a%2520Webpage%2520in%2520Google%2520Chrome_%2520%25E2%2580%2593%2520WebNots.html

Reviewer #2 (Public Review):

The authors devised novel and interesting experiments using high precision human MEG to demonstrate the propagation of beta oscillation events along two axes in the brain. Using careful analysis, they show different properties of beta events pre- and post movement, including changes in amplitude. Due to beta's prominent role in motor system dynamics, these changes are therefore linked to behavior and offer insights into the mechanisms leading to movement. The linking of wave-like phenomena and transient dynamics in the brain offers new insight into two paradigms about neural dynamics, offering new ways to think about each phenomena on its own.

Although there is a substantial, and recent, body of literature supporting the conclusions that beta and other neural oscillations are transient, care must be taken when analyzing the data and the resulting conclusions about beta properties in both time and space. For example, modifying the threshold at which beta events are detected could alter their reported properties and expression in space and time. The authors should therefore performing parameter sweeps on e.g. the thresholds for detection of oscillation bursts to determine whether their conclusions on beta properties and propagation hold. If this additional analysis does not change their story, it would lend confidence in the results/conclusions.

Determining the generators of beta events at different locations is a tricky issue. The authors mentioned a single generator that is responsible for propagating beta along the two axes described. However, it is not clear through what mechanism the beta events could travel along the neural substrate without additional local generators along the way. Previous work on beta events examined how a sequence of synaptic inputs to supra and infragranular layers would contribute to a typical beta event waveform. Although it is possible other mechanisms exist, how might this work as the beta events propagate through space? Some further explanation/investigation on these issues is therefore warranted.

Once the mechanisms have been better understood, a question of how much the results generalize to other oscillation frequencies and other brain areas. On the first question of other oscillation frequencies, the authors could easily test whether nearby frequency bands (alpha and low gamma) have similar properties. This would help to determine whether the observations/conclusions are unique to beta, or more generally applicable to transient bursts/waves in the brain. On the second issue of applicability to other brain areas, the authors could relate their work to transient bursts and waves recorded using ECoG and/or iEEG. Some recent work on traveling waves at the brain-wide level would be relevant for such comparisons.

If the source code could be provided on github along with documentation and a standard "notebook" on use other researchers would benefit greatly.

I found this pretty unclear. A code unit is an encoding-dependent "minimal chunk". For example, in UTF-8, a code unit is a 8 bit chunk; in UTF-16, a code unit is a 16 bit chunk. An individual code unit is generally not interpretable outside its context. A code unit can indeed be described by a "bit sequence", but that misses the "word size" notion.

See https://en.wikipedia.org/wiki/Character_encoding#Terminology.

Reviewer #3 (Public Review):

Leander et al used deep mutational scanning to assess the effect of nearly all possible point mutations on four homologous bacterial allosteric transcription factors (aTFs). In particular, they identified mutations that abrogated the transcription factor response to a small molecule effector. The authors go on to use machine learning to determine which physicochemical properties distinguish mutations with allostery-eliminating effects from those without an effect. They report that mutations that eliminate the allosteric response to small molecules are quite variable across homologs and that global features are more predictive of which mutations will break allostery relative to local properties. Overall, the experimental strategy is well-chosen, and a comprehensive comparison of mutational sensitivity across allosteric homologs is highly important to understand how conserved (or not) the implementation of allostery is across homologs. Moreover, the idea to use machine learning to assess which features are most predictive of "allosteric hotspots" is very nice, and provides some insight into what physical properties distinguish mutations that influence allostery. The authors include some interesting results on transfer learning (evaluating whether models trained on one protein predict allostery in another), and the use of alternate sequence representations (e.g. UniRep) in their machine learning analyses. However - at least in the manuscript's present form - the paper suffers from key conceptual difficulties and a lack of rigor in data analysis that substantially limits one's confidence in the authors' interpretations. More specifically:

1) A key conceptual challenge shaping the interpretation of this work lies in the definition of allostery, and allosteric hotspot. The authors define allosteric mutations as those that abrogate the response of a given aTF to a small molecule effector (inducer). Thus, the results focus on mutations that are "allosterically dead". However, this assay would seem to miss other types of allosteric mutations: for example, mutations that enhance the allosteric response to ligand would not be captured, and neither would mutations that more subtly tune the dynamic range between uninduced ("off) and induced ("on") states (without wholesale breaking the observed allostery). Prior work has even indicated the presence of TetR mutations that reverse the activity of the effector, causing it to act as a co-repressor rather than an inducer (Scholz et al (2004) PMID: 15255892). Because the work focuses only on allosterically dead mutations, it is unclear how the outcome of the experiments would change if a broader (and in our view more complete) definition of allostery were considered.

2) The experimental determination of which mutations impacted allostery is given only a limited description in the methods, but if we understand what was done, the analysis seems to neglect both (1) important caveats due to assay specifics and (2) more general limitations of deep mutational scanning data. In particular:<br /> a. The separation in fluorescence between the uninduced and induced states (the assay dynamic range, or fold induction) varies substantially amongst the four aTF homologs. Most concerningly, the fluorescence distributions for the uninduced and induced populations of the RolR single mutant library overlap almost completely (Figure 1, supplement 1), making it unclear if the authors can truly detect meaningful variation in regulation for this homolog.<br /> b. The methods state that "variants with at least 5 reads in both the presence and absence of ligand in at least two replicates were identified as dead". However, the use of a single threshold (5 reads) to define allosterically dead mutations across all mutations in all four homologs overlooks several important factors:<br /> i. Depending on the starting number of reads for a given mutation in the population (which may differ in orders of magnitude), the observation of 5 reads in the gated non-fluorescent region might be highly significant, or not significant at all. Often this is handled by considering a relative enrichment (say in the induced vs uninduced population) rather than a flat threshold across all variants.<br /> ii. Depending on the noise in the data (as captured in the nucleotide-specific q-scores) and the number of nucleotides changed relative to the WT (anywhere between 1-3 for a given amino acid mutation) one might have more or less chance of observing five reads for a given mutation simply due to sequencing noise.<br /> iii. Depending on the shape and separation of the induced (fluorescent) and uninduced (non-fluorescent) population distributions, one might have more or less chance of observing five reads by chance in the gated non-fluorescent region. The current single threshold does not account for variation in the dynamic range of the assay across homologs.<br /> c. The authors provide a brief written description of the "weighted score" used to define allosteric hotspots (see y-axis for figure 1B), but without an equation, it is not clear what was calculated. Nonetheless, understanding this weighted score seems central to their definition of allosteric hotspots<br /> d. The authors do not provide some of the standard "controls" often used to assess deep mutational scanning data. For example, one might expect that synonymous mutations are not categorized as allosterically dead using their methods (because they should still respond to ligand) and that most nonsense mutations are also not allosterically dead (because they should no longer repress GFP under either condition). In general, it is not clear how the authors validated the assay/confirmed that it is giving the expected results.<br /> 3) In several places, the manuscript lacks important statistical analyses needed to firmly establish the authors' claims<br /> a. The authors performed three replicates of the experiment, but reproducibility across replicates and noise in the assay is not presented/discussed<br /> b. In the analysis of long-range interactions, the authors assert that "hotspot interactions are more likely to be long-range than those of non-hotspots", but this was not accompanied by a statistical test (Figure 2 - figure supplement 1)

4) Data availability and analysis transparency need improvement. The raw fastq reads do not seem to be publicly available, nor did we see access to the code used to perform the analysis. If the code is not provided, the description of the analysis in the methods section needs to be more detailed for reproducibility.

Overall, these concerns with fundamental aspects of the data analysis make it challenging to assess the reproducibility of the results, the fidelity of the assay (in reporting allosterically dead mutations), and the extent to which the data robustly support the authors' claims.

esto sale oiiiiga? y ahoraç???

https://web.archive.org/web/20220901193117/https://interconnected.org/home/2022/09/01/carbon

Matt went to The Conference in Malmo, nice. Videos seem to be online, find them. I've already seen websites that color code their current emissions based on energy source / mix. E.g. Michelle Thorne's when we talked about [[Digirights and climate justice TGL]] last June.

Key phrase is the switch to full solar (plus storage I assume) is a switch to abundance (even if [[Intermittency van infra productie 20200718095613]] still plays a role) away from a sense of scarcity. Vgl [[MakerHouseholds 20100420071013]] wrt household as a productive node in a community and wrt resilience.

what is that ?

interesting

weakness of the monolithic systems

Note: This rebuttal was posted by the corresponding author to Review Commons. Content has not been altered except for formatting.

Learn more at Review Commons

Reply to the reviewers

Reply to the Reviewers

We thank the reviewers for dedicating time to review our manuscript and providing highly valuable feedback. Please find below a point-by-point answer.

Reviewer #1 (Evidence, reproducibility and clarity (Required)):

Summary:

In the present manuscript, van der Plas et al. compellingly illustrated a novel technique for engendering a whole-brain functional connectivity map from single-unit activities sampled through a large-scale neuroimaging technique. With some clever tweaks to the restricted Boltzmann Machine, the cRBM network is able to learn a low-dimensional representation of population activities, without relying on constrained priors found in some traditional methods. Notably, using some 200 hidden layer neurons, the employed model was able to capture the dynamics of over 40,000 simultaneously imaged neurons with a high degree of accuracy. The extracted features both illustrate the anatomical topography/connectivities and capture the temporal dynamics in the evolution of brain states. The illustrated technique has the potential for wide-spread applications spanning diverse recording techniques and animal species. Furthermore, the prospectives of modeling whole-brain network dynamics in 'neural trajectory' space and of generating artificial data in silico make for very enticing reasons to adopt cRBM.

Major comments:

1. Line 164. The authors claim that conventional methods "such as k-means, PCA and non-negative matrix factorization" cannot be quantitatively assessed for quality on the basis that they are unable to generate new artificial data. Though partly true, in most neuroscience applications, this is hardly cause for concern. Most dimensionality reduction methods (with few exceptions such as t-sne) allow new data points to be embedded into the reduced space. As such, quality of encoding can be assessed by cross-validation much in the same way as the authors described, and quantified using traditional metrics such as percentage explained variance. The authors should directly compare the performance of their proposed model against that of NNMF and variational auto-encoders. Doing so would offer a more compelling argument for the advantage of their proposed method over more widely-used methods in neuroscience applications. Furthermore, a direct comparison with rastermap, developed by Stringer lab at Janelia (https://github.com/MouseLand/rastermap), would be a nice addition. This method presents itself as a direct competitor to cRBM. Additionally, the use of GLM doesn't do complete justice to the comparison point used, since a smaller fraction of data were used for calculating performance using GLM, understandably due to its computationally intensive nature.

PLANNED REVISION #2

We thank the reviewer for the comment, and certainly agree that there are multiple methods for unsupervised feature extraction from data and that they can be validated for encoding quality by cross-validation. However, we stress that reconstructing through a low-dimensional, continuous bottleneck is a different (and arguably, easier) task, than generating whole distributions. Reconstruction delineates the manifold of possible configurations, whereas generative modeling must weigh such configurations adequately. Moreover, none of the methodologies mentioned can perform the same tasks as cRBMs. For instance, NNMF learns localized assemblies, but cannot faithfully model inhibitory connections since, by definition, only non-negative weights are learnt. Also, the connection between the learnt assemblies and the underlying connectivity is unclear. Similarly, rastermap is an algorithm for robustly i) sorting neurons along a set number of dimensions (typically 1 or 2) such that neighboring neurons are highly correlated, and ii) performing dimensionality reduction by clustering along these dimensions. Because Rastermap uses k-means as the basis for grouping together neurons, it does not quantify connections between neurons and assemblies, nor assign neurons to multiple assemblies. Moreover, it is not a generative model, and thus cannot predict perturbation experiments, infer connectivities or assign probabilities to configurations. Therefore, we do not believe that NNMF or Rastermap would be a suitable alternative for cRBM in our study. We nonetheless appreciate the reviewer’s suggestions and agree that we should motivate more clearly why these methods are not applicable for our purposes. Therefore, to emphasize the relative merit of cRBM with respect to other unsupervised algorithms, we now provide a table (Supplementary Table 2) that lists their specific characteristics. We stress that we do not claim that cRBM are consistently better than these classical tools for dimensionality reduction, but focus only on the properties relevant to our study.

Further, we agree that VAEs, which also jointly learn a representation and distribution of data, are close competitors of cRBMs. In Tubiana et al. Neural Computation 2019, we previously compared sparse VAEs with cRBMs for protein sequence modeling, and found that RBMs consistently outperformed VAEs in terms of the interpretability-performance trade-off. In the revised manuscript, we propose to repeat the comparison with VAE for zebrafish neural recordings, and expect similar conclusions.

As for GLM, it is true that the comparison involved subsampling of the neurons (due to the very high computational cost of GLM, where we could estimate the connectivity of 1000 neurons per day). This was already denoted in the relevant figure caption, as the reviewer has seen, but we have now also clarified this point in Methods 7.10.3. Still, we performed our GLM analysis on 5000 neurons (using all neurons as regressors), which is 10% of all neurons, and we believe this is a sufficient number for comparison. This emphasizes the ability of our optimized cRBMs to handle very large datasets, such as the presently used zebrafish whole brain recordings.

Line 26. The authors describe their model architecture as a formalization of cell assemblies. Cell assemblies, as originally formulated by Hebb, pertains to a set of neurons whose connectivity matrix is neither necessarily complete nor symmetric. Critically, in the physiological brain, the interactions between the individual neurons that are part of an assembly would occur over multiple orders of dependencies. In a restricted Boltzmann machine, neurons are not connected within the same layer. Instead, visible layer neurons are grouped into "assemblies" indirectly via a shared connection with a hidden layer neuron. Furthermore, a symmetrical weight matrix connects the bipartite graph, where no recurrent connectivities are made. As such, the proposed model still only elaborates symmetric connections pertaining to first-order interactions (as illustrated in Figure 4C). Such a network may not be likened with the concept of cell assemblies. The authors should refrain from detailing this analogy (of which there are multiple instances of throughout the text). It is true that many authors today refer to cell assemblies as any set of temporally-correlated neurons. However, saying "something could be a cell assembly" is not the same as saying "something is a cell assembly". How about sticking with cRBM-based cell assemblies (as used in section 2.3) and defining it beforehand?

We thank the reviewer for this excellent question. We agree that there is, in general, a discrepancy between computationally-defined assemblies and conceptual/neurophysiological definition of cell assemblies. We have added a clarification in Results 2.1 to clarify the use of this term when it first occurs in Results. However, we still believe that our work contributes to narrowing the gap. Indeed, our RBM-defined assemblies are i) localized, ii) overlapping, iii) rooted in connectivity patterns (both excitatory and inhibitory), and iv) cannot be reduced to a simple partitioning of the brain with full & uniform connectivity within and between partitions. This is unlike previous work based on clustering (no overlaps or heterogeneous weights), NNMF (no inhibition) or correlation network analysis (no low-dimensional representation).

Regarding the specific comments pointed here, we stress that:

• Effective interactions between neurons are not purely pairwise (“First order”), due to the usage of the non-quadratic potential. (see Eqn 12-13). If the reviewer means by “First-order” interactions the lack of hierarchical organization, we agree, to some extent: in the current formulation, correlations between assemblies are mediated by overlaps between their weights. Fully-hierarchical organization, e.g. by using Deep Boltzmann Machines or pairwise connections within the hidden layer is an interesting future direction, but on the other hand may make it hard to clearly identify assemblies as they might be spread out over multiple layers
• Neurons that participate in a given assembly (as defined by a specific hidden unit) are not all connected with one another with equal strength. Indeed, these neurons may participate in other assemblies, resulting in heterogeneity of connections (see Eqn. 15-17) and interactions between assemblies.
• We acknowledge that the constraint of symmetrical connections is a core limitation of our method. Arguably, asymmetric connections are critical for predicting temporal evolution but less important for inferring a steady-state distribution from data, as we do here. In the revised submission, we added a new paragraph in the discussion section (lines 351-357) in which these limitations are discussed, including the imposed symmetry of the connections and the lack of hierarchical structures, copied below. We trust that this addresses the reviewer’s criticism:

In sum, cRBM-inferred cell assemblies display many properties that one expects from physiological cell assemblies: they are anatomically localized, can overlap, encompass functionally identified neuronal circuits and underpin the collective neural dynamics (Harris, 2005, 2012; Eichenbaum, 2018). Yet, the cRBM bipartite architecture lacks many of the traits of neurophysiological circuits. In particular, cRBMs lack direct neuron-to-neuron connections, asymmetry in the connectivity weights and a hierarchical organization of functional dependencies beyond one hidden layer. Therefore, to what extent cRBM-inferred assemblies identify to neurophysiological cell assemblies, as postulated by Hebb (1949) and others, remains an open question.

I would strongly recommend adding a paragraph discussing the limitation of using the cRBM, things future researchers need to keep in mind before using this method. One such recommendation is moving the runtime-related discussion for cRBM, i.e. 8-12 hrs using 16 CPU from Methods to Discussion, since it's relevant for an algorithm like this. Additionally, a statement mentioning how this runtime will increase with the length of recordings and/or with the number of neurons might be helpful. What if the recordings were an hour-long rather than 25mins. This would help readers decide if they can easily use a method like this.

We thank the reviewer for the suggestion, and agree that it is important to cover the computational cost in the main text. Regarding the runtime for longer recordings, the general rule of thumb is that the model requires a fixed number of gradient updates to converge (20-80k depending on the data dimensionality) rather than a fixed number of epochs. Thus, runtime should not depend on recording length, as the number of epochs can be reduced for longer recordings. While we did not verify this rule for neural recordings, this is what we previously observed when modeling protein/DNA sequence data sets, whose size range from few hundreds to hundreds of thousands of samples (Tubiana et al., 2019, eLife; Tubiana et al. 2019, Neural Computation; Bravi et al. Cell Systems 2021; Bravi et al. PLOS CB 2021; Fernandez de Cossio Diaz et al. Arxiv 2022 Di Gioacchino et al. BiorXiv 2022). We have now added a summary of these points in Methods 7.7.2, also refer to this with explicit mention of the runtime in the Discussion, end of 2nd paragraph:

By implementing various algorithmic optimizations (Methods 7.7), cRBM models converged in approximately 8-12 hours on high-end desktop computers (also see Methods 7.7.2).

Line 515. A core feature of the proposed compositional RBM is the addition of a soft sparsity penalty over the weight matrix in the likelihood function. The authors claim that "directed graphical models" are limited by the a priori constraints that they impose on the data structure. Meanwhile, a more accurate statistical solution can be obtained using a RBM-based model, as outlined by the maximum entropy principle. The problem with this argument is that the maximum entropy principle no longer applies to the proposed model with the addition of the penalty term. In fact, the lambda regularization term, which was estimated from a set of data statistics motivated by the experimenter's research goals (Figure S1), serves to constrict the prior probability. Moreover, in Figure S1F, we clearly see that reconstruction quality suffers with a higher penalty, suggesting that the principle had indeed been violated. That being said, RBMs are notoriously hard to train, possibly due to the unconstrained nature of the optimization. I believe that cRBM can help bring RBM into wider practical applications. The authors could test their model on a few values of the free parameter and report this as a supplementary. I believe that different parameters of lambda could elaborate on different anatomical clusters and temporal dynamics. Readers who would like to implement this method for their own analysis would also benefit tremendously from an understanding of the effects of lambda on the interpretation of their data. Item (1) on line 35 (and other instances throughout the text) should be corrected to reflect that cRBM replaces the hard constraints found in many popular methods with a soft penalty term, which allows for more accurate statistical models to be obtained.

We thank the reviewer for their analysis and suggestion. Indeed, adding the regularization term - not present in the classical formulation of the RBM (Hinton & Salakhutdinov, 2006, Science) - was critical for significantly enhancing its performance, which allowed us to implement this model on our large scale datasets (~50K visible units). We agree that providing more information on the effect of the regularization term will benefit readers who would like to use this method, and we propose to add this in the revised manuscript, which would implement the reviewer’s suggestion. See “PLANNED REVISION #1”.

The reviewer’s comment on the Maximum Entropy issue calls for some clarification. The maximum entropy principle is a recipe for finding the least constrained model that reproduces specified data-dependent moments. However, it cannot determine which moments are statistically meaningful in a finite-sized data set. A general practice is to only include low-order moments (1st and 2nd), but this is sometimes already too much for biological data. Regularization provides a practical means to select stable moments to be fitted and others to be ignored. This can be seen from the optimality condition, which writes, e.g., for the weights wi,mu:

| i h,mu>data - i h,mu>model | i,mu = 0.

i h,mu>data - i h,mu>model | = lambda sign(wi,mu) if |wi,mu| > 0.

Essentially, this lets the training decide which subset of the constraints should actually be used. Thus, regularized models are closer to the uniform distribution (g=w=0), and actually have higher entropy than unregularized one (see, e.g., Fanthomme et al. Journal of Statistical Mechanics, 2022). Therefore, we believe that a regularized maximum entropy model can still be considered a bona fide MaxEnt model. This formulation should not be confused with another formulation (that perhaps the reviewer has in mind) where a weighted sum of the entropy and the regularization term is maximized under the same moment-matching constraints. In this case, we agree that maximum entropy principle (MaxEnt) would be violated.

The choice of regularization value should be dictated by bias-variance trade-off considerations. Ideally, we would use the same criterion as for training, i.e., maximization of log-likelihood for the held-out test set, but it is intractable. Thus, we used a consensus between several tractable performance metrics as a surrogate; we believe this consensus to be principally independent of the research goal. While the reconstruction error indeed increases for large regularization values, this is simply because too few constraints are retained at high regularizations.

Essentially, the parameters selected by likelihood maximization find the finest assembly scale that can be accommodated by the data presented. Thus, the number and size of the assemblies are not specified by the complexity of the data set alone. Rather, the temporal resolution and length of the recordings play a key role; higher resolution recordings will allow the inference of a larger number of smaller assemblies, and enable the study of their hierarchical organization.

That being said, we fully agree that the regularization strength and number of hidden units have a strong impact on the nature of the representation learnt. In the revised manuscript, we will follow the reviewer’s suggestion and provide additional insights on the effect of these parameters on the representation learnt (please see revision plan).

Minor comments:

From a neuroscience point of view, it might be interesting to show what results are achieved using different values of M (say 100 or 300), rather than M=200, while still maintaining the compositional phase. Is there any similarity between the cRBM-based cell assemblies generated at different values of M? Is there a higher chance of capturing certain dynamics either functional or structural using cRBM? For example, did certain cRBM-based cell assemblies pop up more frequently than others at all values of M (100,200,300)?

This point will be addressed in the future, as detailed in our response to reviewer 2 (see PLANNED REVISION #1).

The authors have mentioned that this approach can be readily applied to data obtained in other animal models and using different recording techniques. It might be nice to see a demonstration of that.

We agree that showing additional data analysis would be interesting, but we feel that it would overburden the supplementary section of the manuscript, which is already lengthy. In previous works, we and collaborators have used cRBMs for analyzing MNIST data (Tubiana & Monasson, 2017, PRL; Roussel et al. 2022 PRE), protein sequence data (Tubiana et al., 2019, eLife; Tubiana et al. 2019, Neural Computation; Bravi et al. Cell Systems 2021; Bravi et al. PLOS CB 2021; Fernandez de Cossio Diaz et al. Arxiv 2022), DNA sequences (Di Gioacchino et al. BiorXiv 2022), spin systems (Harsh et al. J. Phys. A 2020), etc. Many are included as example notebooks - next to the zebrafish data - in the linked code repository. For neural data, we have recently shared our code with another research group working on mice auditory cortex (2-photon, few thousands of neurons, Léger & Bourdieu). Preliminary results are encouraging, but not ready for publication yet.

Line 237. The justification for employing a dReLU transfer function as opposed to ReLU is unclear, at least within the context of neurobiology. Given that this gives rise to a bimodal distribution for the activity of HUs, the rationale should be clearly outlined to facilitate interpretability.

We thank the reviewer for the question. As we detail in the manuscript (Methods), the dReLU potential is one of the sufficient requirements for the RBM to achieve the compositional phase. The compositional phase is characterized by localized assemblies that co-activate to generate the whole-brain neural dynamics. This property reflects neurobiological systems (Harris, 2005, Neuron), which is one of the reasons why we employed compositional RBMs for this study.

As the reviewer points out, the HUs that we infer exhibit bimodal activity (Figure 4). Importantly, the HU activity is not constrained by the model to take this shape, as dReLU potentials allow for several activity distributions (see Methods 7.5.4; “Choice of HU potential”). In fact, ReLU potentials are a special case of dReLU (by $(\gamma_{\mu, -} \to \infty)$), so our model allows HU potentials to behave like ReLUs, but in practice they converge to a double-well potential for almost all HUs, leading to bimodal activity distributions.

Following the suggestion of the reviewer, we have now added this detail for clarity in Methods 7.5.4 and referenced this Methods section at line 237.

Reviewer #1 (Significance (Required)):

van der Plas et al. highlighted a novel dimensionality reduction technique that can be used with success for discerning functional connectivities in large-scale single-unit recordings. The proposed model belongs to a large collection of dimensionality reduction techniques (for review, Cunningham, J., Yu, B. Dimensionality reduction for large-scale neural recordings. Nat Neurosci 17, 1500-1509 (2014). https://doi.org/10.1038/nn.3776; Paninski, L., & Cunningham, J. P. (2018). Neural data science: accelerating the experiment-analysis-theory cycle in large-scale neuroscience. Current opinion in neurobiology, 50, 232-241.). The authors themselves highlighted some of the key methods, such as PCA, ICA, NNMF, variational auto-encoders, etc. The proposed cRBM model has also been published a few times by the same authors in previous works, although specifically pertaining to protein sequences. The use of RBM-like methods in uncovering functional connectivities is not novel either (see Hjelm RD, Calhoun VD, Salakhutdinov R, Allen EA, Adali T, Plis SM. Restricted Boltzmann machines for neuroimaging: an application in identifying intrinsic networks. Neuroimage. 2014 Aug 1;96:245-60. doi: 10.1016/j.neuroimage.2014.03.048.). However, given that the authors make a substantial improvement on the RBM network and have demonstrated the value of their model using physiological data, I believe that this paper would present itself as an attractive alternative to all readers who are seeking better solutions to interpret their data. However, as I mentioned in my comments, I would like to see more definitive evidence that the proposed solution has a serious advantage over other equivalent methods.

Reviewer's expertise:

This review was conducted jointly by three researchers whose combined expertise includes single-unit electrophysiology and two-photon calcium imaging, using which our lab studies the neurobiology of learning and memory and spatial navigation. We also have extensive experience in computational neuroscience, artificial neural network models, and machine learning methods for the analysis of neurobiological data. We are however limited in our knowledge of mathematics and engineering principles. Therefore, our combined expertise is insufficient to evaluate the correctness of the mathematical developments.

Reviewer #2 (Evidence, reproducibility and clarity (Required)):

In their manuscript, van der Plas et al. present a generative model of neuron-assembly interaction. The model is a restricted Boltzmann machine with its visible units corresponding to neurons and hidden units to neural assemblies. After fitting their model to whole-brain neural activity data from larval zebrafish, the authors demonstrate that their model is able to replicate several data statistics. In particular, it was able to replicate the pairwise correlations between neurons as well as assemblies that it was not trained on. Moreover, the model allows the authors to extract neural assemblies that govern the population activity and compose functional circuits and can be assigned to anatomical structures. Finally, the authors construct functional connectivity maps from their model that are then shown to correlate with established structural connectivity maps.

Overall, the authors present convincing evidence for their claims. Furthermore, the authors state that their code to train their restricted Boltzmann machine models is already available on GitHub and that the data underlying the results presented in this manuscript will be made publicly available upon publication, which will allow people to reproduce the results and apply the methods to their data.

One thing the authors could maybe discuss a bit more is the "right" parameter value M, especially since they used the optimal value of 200 found for one sample also for all the others. More specifically, how sensitive are the results to this value?

PLANNED REVISION #1

In the following we jointly address three of the reviewers’ questions (2 from reviewer 1, and 1 from reviewer 2).

Shortly summarized, the cRBM model has 2 free parameters; the number of hidden units M and the regularization parameter lambda. In figures 2 and S1 we optimize their values through cross-validation, and then perform our further analyses on models with these optimal values. The reviewers ask us to examine the outcome of the model for slightly different values of both parameters, in particular in relation to the sensitivity of the cRBM results to selecting the optimal parameters and the change in inferred assemblies and their dynamics.

We thank the reviewers for these questions and appreciate their curiosity to understand the effects of changing either of these two free parameters.

We inspected these models when we performed the model selection (of Figures 2 and S1), but did not formalize our findings into figures for the manuscript. We found that small changes in the parameter setting did not abruptly change the inferred assemblies (e.g., M=100) apart from slightly changing in size, so we expect that the statistics that we intend to include in the proposed supplementary figure would reflect that, and it would definitely benefit the manuscript to include this analysis. Very-low-M settings are interesting to include, because assemblies are much larger - essentially merging smaller assemblies of higher-M models - at the cost of model performance.

We propose to create additional supplementary figures that address these questions. As suggested, we will pick a few example cRBMs with different parameter settings (below-optimal M, above-optimal M, and same for lambda), as well as very low M settings (M~20 or 50). We will then show example assemblies and assembly dynamics, as well as the relevant statistics (assembly size, dynamics time scale etc) that describe them.

And, what happens if one would successively increase that number, would the number of assemblies (in the sense of hidden units that strongly couple to some of the visible units) eventually saturate?

This point will be addressed by inspecting models at different M values (see Revision Plan #1). We would like to further answer this question by referring to past work. In Tubiana et al., 2019, elife (Appendix 1) we have done this analysis, and the result is consistent with the reviewer’s intuition. Because of the sparsity regularization, if M becomes larger than its optimum, the assemblies further sparsify without benefiting model performance, and eventually new assemblies duplicate previous assemblies or become totally sparse (i.e., all weights = 0) to not further induce a sparsity penalty in the loss function. So the ‘effective’ number of assemblies indeed saturates for high M.

Moreover, regarding the presentation, I have a few minor suggestions and comments that the authors also might want to consider:

* In Figure 6C, instead of logarithmic axes, it might be better to put the logarithmic connectivity on a linear axis. This way the axes can be directly related to the colour bars in Figures 6A and B.

We agree and thank you for the suggestion. We have changed this accordingly (and also in the equivalent plots in figure S6).

* In Equation (8), instead of $\Gamma_{\mu}(I)$ it should be $\Gamma_{\mu}(I_{\mu}(v))$.

Done, thank you.

* In Section 7.0.5, it might make sense to have the subsection about the marginal distributions before the ones about the conditional distributions. The reason would be that if one wants to confirm Equation (8) one necessarily has to compute the marginal distribution in Equation (12) first.

We thank the reviewer for the suggestion, but respectfully propose to leave the section ordering as is. We understand what the reviewer means, but Equation (8) can also be obtained by factorizing P(v,h) Equation (7) and removing the v_i dependency. In Equation (8), \Gamma can then be obtained by normalization. We believe this flow aligns better with the main text (where conditionals come first, when used for sampling, followed by the marginal of P(v) used for the functional connectivity inference).

* In Line 647f, the operation the authors are referring to is strictly speaking not an L1-norm of the matrix block. It might be better to refer to that e.g. as a normalised L1-norm of the matrix block elements.

Done, thank you.

* In Line 22, when mentioning dimensionality reduction methods to identify assemblies, it might make sense to also reference the work by Lopes-dos-Santos et al. (2013, J. Neurosci. Methods 220).

Done, thank you for the suggestion.

Reviewer #2 (Significance (Required)):

The work presented in this manuscript is very interesting for two reasons. First, it has long been suggested that assemblies are a fundamental part of neural activity and this work seems to support that by showing that one can generate realistic whole-brain population activity imposing underlying assembly dynamics. Second, in recent years much work has been devoted to developing methods to find and extract neural assemblies from data and this work and the modelling approach can also be seen as a new method to achieve that. As such, I believe this work is relevant for anyone interested in neural population activity and specifically neural assemblies and certainly merits publication.

Regarding my field of expertise, I used to work on data analysis of neural population activity and in particular on the question of how one can extract neural assemblies from data. I have to say that I have not much experience with fitting statistical models to data, so I can't provide any in-depth comments on that part of the work, although what has been done seems plausible.

Reviewer #3 (Evidence, reproducibility and clarity (Required)):

Summary: Understanding the organization of neural population activities in the brain is one of the most important questions in neuroscience. Recent technique advance has enabled researchers to record a large number of neurons and some times the whole brain. Interpreting and extracting meaningful insights from such data sets are challenging. van der Pals \textit{et al} applied a generative model called compositional Restricted Boltzmann Machine (cRBM) to discover neuron assemblies from spontaneous activities of zebra fish brain. They found that neurons can be grouped into around 200 assemblies. Many of them have clear neurophysiological meaning, for example, they are anatomically localized and overlapped with known neural circuits. The authors also inferred a coarse-grained functional connectivity which is similar to known structural connectivity.

The structure of the paper is well organized, the conclusion seems well supported by their numerical results. While this study provides a compelling demonstration that cRBM can be used to uncover meaningful structures from large neural recordings, the following issues limit my enthusiasm.

Major:

1) The overall implication is not clear to me. Although the authors mentioned this briefly in the discussion. It is not clear what else do we learn from discovered assemblies beyond stating that they are consist with previous study. For example, the author could have more analysis of the assembly dynamics, such as whether there are low dimensional structure etc.

First, we will comment on our analysis of the assemblies, before we continue to discuss the main implications of our work, which we believe are the inferred generative model of the zebrafish brain and the perturbation-based connectivity matrix that we discovered. Further, we have implemented the reviewer’s suggestion of analyzing the low-dimensional structure of the hidden unit activity, as further detailed in Question 7.

Indeed, the example assemblies that we show in Figure 3 have been thoroughly characterized in previous studies, which is why we chose to showcase these examples. Previous studies (including our own) typically focused on particular behaviors or sensory modalities, and aimed at identifying the involved neural circuit. Here, we demonstrate that by using cRBM on spontaneous activity recordings, one can simultaneously identify many of those circuits. In other words, these functional circuits/assemblies activate spontaneously, but in many different combinations and perhaps infrequently, so that it is very difficult to infer them from the full neural dynamics that they generate. cRBM has been able to do so, and Figure 3 (and supplementary video 1) serve to illustrate the variety of (known) circuits and assemblies that it inferred, some of which may represent true but not yet characterized circuits, which thus provide hypotheses for subsequent studies.

Further, we believe that the implication of our study goes beyond the properties of the assemblies we’ve identified, in several ways.

We demonstrate the power of cRBM’s generative capacity for inferring low-dimensional representations in neuroscience. Unlike standard dimensional reductionality methods, generative models can be assessed by comparing the statistics of experimental vs in-silico generated data. This is a powerful approach to validate a model, rarely used in neuroscience because of the scarcity of generative models compatible with large-scale data, and we hope that our study will inspire the use of this method in the field. We have made our cRBM code available, including notebook tutorials, to facilitate this.

The generative aspect of our model allowed us to predict the effect of single-neuron perturbations between all ~ 10^9 pairs of neurons per fish, resulting in a functional connectivity matrix.

We believe that the functional connectivity matrix is a major result for the field, similar to the structural connectivity matrix from Kunst et al., 2019, Neuron. The relation between functional and structural connectivity is unknown and of strong interest to the community (e.g., Das & Fiete, 2020, Nature Neuroscience). Our results allowed for a direct comparison of whole-brain region-by-region structural and functional connectivity. We were thus able to quantify the similarity between these two maps, and to identify specific region-pair matches and non-matches of functional and structural connectivity - which will be of particular interest to the zebrafish neuroscience community for developing future research questions.

Further, using these trained models - that will be made public upon publication - anyone can perform any type of in silico perturbation experiments, or generate endless artificial data with matching data statistics to the in vivo neural recordings.

We hope that this may convince the reviewer of the multiple directions of impact of our study. We will further address their comment on analysis of assembly dynamics below (question 7).

2) The learning algorithm of cRBM can be interpreted as matching certain statistics between the model and the experiment. For a general audience, it is not easy to understand $\langle h_{\mu} \rangle_{data}, \langle v_ih_{\mu}\rangle_{data}$. Since these are not directly calculated from experimental observed activities $v_i$, but rather the average is conditioned on the empirical distribution of $p(v_i)$. For example, the meaning of $\langle v_ih_{\mu}\rangle_{data}$ means

\begin{equation}

\langle v_ih_{\mu}\rangle_{data} = \frac{1}{l}\sum_{\mathbf{v}\in S} \mathbb{E}{p(\mathbf{h|v})}(v_ih{\mu}),

\end{equation}

where $S$ is the set of all observed neural activities: $S = {\mathbf{v}^1, \cdots, \mathbf{v}^l}$. The authors should explain this in the main text or method, since they are heavily loaded in figure 2.

We thank the reviewer for their suggestion, and have now implemented this. Their mathematics are correct; and we agree that it is not easy to understand without going through the full derivation of the (c)RBM. At the same time, we have tried not to alienate readers who might be more interested in the neuroscience findings than in understanding the computational method used. Therefore, we have kept mathematical details in the main text to a minimum (and have used schematics to indicate the statistics in Figures 2C-G), while explaining it in detail in Methods.

Accordingly, we have now extended section 7.10.2 (“Assessment of data statistics”) that explains how the data statistics were computed in Methods (and have referenced this in Results and in Methods 7.5.5), using the fact that we already explain the process of conditioning on __v __in Methods 7.5.1. The following sentences were added:

[..] However, because (c)RBM learn to match data statistics to model statistics (see Methods

7.5.5), we can directly compare these to assess model performance. [..]

[..]

For each statistic 〈 fk〉 we computed its value based on empirical data 〈 fk〉_data and on the model 〈 fk〉_model, which we then quantitatively compared to assess model performance. Data statistics 〈 fk〉_data were calculated on withheld test data (30% of recording). Naturally, the neural recordings consisted only of neural data v and not of HU data h. We therefore computed the expected value of __ht at each time point t conditioned on the empirical data _v_t, as further detailed in Methods 7.5.1.

[..]

3) As a modeling paper, it would be great to have some testable predictions.

We thank the reviewer for the enthusiasm and suggestion. We agree, and that is why we have included this in the form of functional connectivity matrices in Figures 5 and 6. To achieve this, we leveraged the generative aspect of the cRBM to perform in silico single-neuron perturbation experiments, which we aggregated to connectivity matrices. In other words, we have used our model to predict the functional connectivity between brain regions using the influence of single-neuron perturbations.

Obtaining a measure of functional connectivity/influence using single-neuron perturbations is also possible using state-of-the-art neuro-imaging experiments (e.g., Chettih & Harvey, 2019, Nature), though not at the scale of our in silico experiments. We therefore verify our predictions using structural data from Kunst et al., 2019, which we have extended substantially. We provide our functional connectivity result in full, and hope that this can inspire future zebrafish research by predicting which regions are functionally connected, which includes many pairs of regions that have not yet directly been studied in vivo.

Minor:

1) The assembly is defined by the neurons that are strongly connected with a given hidden unit. Thus, some neurons may enter different assemblies. A statistics of such overlap would be helpful. For example, a ven diagram in figure 1 that shows how many of them assigned to 1, 2, etc assemblies.

We thank the reviewer for this excellent suggestion. Indeed, neurons can be embedded in multiple assemblies. This is an important property of cRBMs, which deserves to be quantified in the manuscript. We have now added this analysis as a new supplementary figure 4. Neurons are embedded in an assembly if their connecting weight w_{i, \mu} is ‘significantly’ non-zero, depending on what threshold one uses. We have therefore shown this statistic for 3 values of the threshold (0.001, 0.01 and 0.1) - demonstrating that most neurons are strongly embedded in at least 1 assembly and that many neurons connect to more than 1 assembly.

Updated text in Results:

Further, we quantified the number of assemblies that each neuron was embedded in, which showed that increasing the embedding threshold did not notably affect the fraction of neurons embedded in at least 1 assembly (93% to 94%, see Figure S4).

2) What does the link between hidden units in Figure 1B right panel mean?

Thank you for the question, and we apologize for the confusion: if we understand the question right, the reviewer asks why the colored circles under the title ‘Neuronal assemblies of Hidden Units’ are linked. This schematic shows the same network of neurons as shown in gray at the left side of Fig 1B, but now colored by the assembly ‘membership’ of each neuron. Hence, the circles shown are still neurons (and not HUs), and their links still represent synaptic connections between neurons. We apologize for the confusion, and have updated the caption of Fig 1B to explain this better:

“[..] The neurons that connect to a given HU (and thus belong to the associated assembly), are depicted by the corresponding color labeling (right panel).[..]”.

3) A side-by-side comparison of neural activity predicted by model and the experimentally recorded activities would help the readers to appreciate the performance of the model. Such comparison can be done at both single neuron level or assembly level.

We thank the reviewer for this suggestion. The cRBM model is a statistical model, meaning that it fits the statistics of the data, and not the dynamics. The data that it generates therefore (should) adhere to the statistics of the training data, but does not reflect their dynamics. We believe that showing generated activity side-by-side of empirical activity is therefore not a meaningful example of generated data, as this would exemplify the dynamics, which this model is not designed to capture. Instead, in Figure 2, we show the statistics of the generated data versus the statistics of the empirical data (e.g., Fig 2C for the mean activity of all neurons). We believe that this is a better example representation of the generative performance of the model.

4) Definition of reconstruction quality in line 130.

We thank the reviewer for the suggestion, and have added the following sentence after line 130:

The reconstruction quality is defined as the log-likelihood of reconstructed neural data v___{recon} (i.e., __v that is first transformed to the low-dimensional h, and then back again to the high-dimensional __v___{recon}, see Methods 7.10.2).

Further, please note that Methods describes the definition in detail (Eq 18 of the submitted manuscript), although we agree with the reviewer that more detail was required in the Results section at line 130.

5) Line 165. If PCA is compared with cRBM, why other dimensionality reduction methods, such as k-means and non-negative matrix factorization, can not be compared in terms of the sparsity?

Please see answer to question 1 from R1 and Revision Plan #2.

6) Line 260, please provide minimum information about how the functional connectivity is defined based on assemblies discovered by cRBM.

We apologize if this was not clear. The first paragraph of this section (lines 248-259) of the submitted manuscript, provides the detail that the reviewer asks for, and we realize that the sentence of line 260 is better placed in the first paragraph, as it has come across as a very minimal explanation of how functional connectivity is defined.

We have now moved this sentence to the preceding paragraph, as well as specified the Method references (as suggested by this reviewer below), for additional clarity. We thank the reviewer for pointing out this sentence.

7) Some analysis of the hidden units population activities. Such as whether or not there are interesting low dimensional structure from figure 4A.

We thank the reviewer for their suggestion. In our manuscript we have used the cRBM model to create a low-dimensional (M=200) representation of zebrafish neural recordings (N=50,000). The richness of this model owes to possible overlaps between HUs/assemblies that can result in significant correlation in their activities. The latter is illustrated in Figure 4A-C: the activity of some HUs can be strongly correlated.

The reviewer’s suggestion is similar; to perform some form of dimensionality reduction on the low-dimensional HU activity shown in Fig 4. We have now added a PCA analysis to Figure 4 to quantify the degree of low-dimensional structure in the HU dynamics, and show the results in a new panel Figure 4D.

The following text has been added to the Results section:

These clusters of HUs with strongly correlated activity suggest that much of the HU variance could be captured using only a small number of variables. We quantified this by performing PCA on the HU dynamics, finding that indeed 52% of the variance was captured by the first 3 PCs, and 85% by the first 20 PCs (Figure 4D).

We believe that further visualization of these results, such as plotting the PC trajectories, would not further benefit the manuscript. The manuscript focuses on cRBM, and the assemblies/HUs it infers. Unlike PCA, these are not ranked/quantified by how much variance they explain individually, but rather they together ‘compose’ the entire system and explain its (co)variance (Figure 2). Breaking up a dominant activity mode (as found by PCA), such as the ARTR dynamics, into multiple HUs/assemblies, allows for some variation in activity of individual parts of the ARTR circuit (such as tail movement and eye movement generation), even though at most times the activity of these HUs is coordinated. We hope the reviewer agrees with our motivation to keep the manuscript focused on the nature of cRBM-inferred HUs.

8) Figure 4B right panel, how did the authors annotate the cluster manually? As certain assembly may overlap with several different brain regions, for example, figure 4D.

We thank the reviewer for this question, and we presume they meant to reference figure 3D as an example? For figure 4, as well as Figure 3, we used the ZBrain Atlas (Randlett et al., 2015) for definition of brain regions. This atlas presents a hierarchy of brain regions: for example, many brain regions are part of the rhombencephalon/hindbrain. This is what we used for midbrain/hindbrain/diencephalon. Further, many assemblies are solely confined to Optic Tectum (see Fig 3L), which we therefore used (split by hemisphere). Then, many brain regions are (partly) connected to the ARTR circuit, such as the example assembly of Figure 3D that the reviewer mentions. These we have all labeled as ARTR (left or right), though technically only part of their assembly is the ARTR. These two clusters therefore rather mean ‘ARTR-related’, in particular because their activity is locked to the rhythm of the ARTR (see Fig 4A). The final category is ‘miscellaneous’ (like Figure 3G).

However we agree that this wasn’t clear from the manuscript text, so we have changed the figure 4C caption to mention that ‘ARTR’ stands for ARTR-related assemblies, which we hope clarifies that ARTR-clustered assemblies can exist of multiple, disjoint groups of neurons, which relate to the ARTR circuit.

9) Better reference of the methods cited in the main text. The method part is quite long, it would be helpful to cite the section number when referring it in the main text.

We thank the reviewer for this helpful suggestion, we agree that it would benefit the manuscript to reference specific sections of the Methods. We have now changed all references to Methods to incorporate this.

10) Some discussion about the limitation of cRBM would be great.

We thank the reviewer for this suggestion, and have now included this. As Reviewer 1 had the same suggestion, we refer our answer to questions 2 and 3 from R1 for more detail.

Reviewer #3 (Significance (Required)):

This work provides a timely new technique to extract meaningful neural assemblies from large scale recordings. This study should be interested to both researchers doing either experiments and computation/theory. I am a computational neuroscientist.

Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

Learn more at Review Commons

Referee #2

Evidence, reproducibility and clarity

In their manuscript, van der Plas et al. present a generative model of neuron-assembly interaction. The model is a restricted Boltzmann machine with its visible units corresponding to neurons and hidden units to neural assemblies. After fitting their model to whole-brain neural activity data from larval zebrafish, the authors demonstrate that their model is able to replicate several data statistics. In particular, it was able to replicate the pairwise correlations between neurons as well as assemblies that it was not trained on. Moreover, the model allows the authors to extract neural assemblies that govern the population activity and compose functional circuits and can be assigned to anatomical structures. Finally, the authors construct functional connectivity maps from their model that are then shown to correlate with established structural connectivity maps.

Overall, the authors present convincing evidence for their claims. Furthermore, the authors state that their code to train their restricted Boltzmann machine models is already available on GitHub and that the data underlying the results presented in this manuscript will be made publicly available upon publication, which will allow people to reproduce the results and apply the methods to their data.

One thing the authors could maybe discuss a bit more is the "right" parameter value M, especially since they used the optimal value of 200 found for one sample also for all the others. More specifically, how sensitive are the results to this value? And, what happens if one would successively increase that number, would the number of assemblies (in the sense of hidden units that strongly couple to some of the visible units) eventually saturate?

Moreover, regarding the presentation, I have a few minor suggestions and comments that the authors also might want to consider: - In Figure 6C, instead of logarithmic axes, it might be better to put the logarithmic connectivity on a linear axis. This way the axes can be directly related to the colour bars in Figures 6A and B. - In Equation (8), instead of $\Gamma_{\mu}(I)$ it should be $\Gamma_{\mu}(I_{\mu}(v))$. - In Section 7.0.5, it might make sense to have the subsection about the marginal distributions before the ones about the conditional distributions. The reason would be that if one wants to confirm Equation (8) one necessarily has to compute the marginal distribution in Equation (12) first. - In Line 647f, the operation the authors are referring to is strictly speaking not an L1-norm of the matrix block. It might be better to refer to that e.g. as a normalised L1-norm of the matrix block elements. - In Line 22, when mentioning dimensionality reduction methods to identify assemblies, it might make sense to also reference the work by Lopes-dos-Santos et al. (2013, J. Neurosci. Methods 220).

Significance

The work presented in this manuscript is very interesting for two reasons. First, it has long been suggested that assemblies are a fundamental part of neural activity and this work seems to support that by showing that one can generate realistic whole-brain population activity imposing underlying assembly dynamics. Second, in recent years much work has been devoted to developing methods to find and extract neural assemblies from data and this work and the modelling approach can also be seen as a new method to achieve that. As such, I believe this work is relevant for anyone interested in neural population activity and specifically neural assemblies and certainly merits publication.

Regarding my field of expertise, I used to work on data analysis of neural population activity and in particular on the question of how one can extract neural assemblies from data. I have to say that I have not much experience with fitting statistical models to data, so I can't provide any in-depth comments on that part of the work, although what has been done seems plausible.

Code block below says load 'both packages', it would be useful know that you intend to load AER per the solution, the text only mentions MASS.

Obnoxious.

As someone recently pointed out on HN, it's very common nowadays to encounter the no-one-knows-else-what-they're-doing-here refrain as cover—I don't have to feel insecure about not understanding this because not only am I not alone, nobody else understands it either.

Secondly, if your code is hard to understand regarding its use of this, then your code his hard to understand. this isn't super easy, but it's also not hard. Your code (or the code you're being made to wade into) probably just sucks. The this confusion is making you confront it, though, instead of letting it otherwise fly under the radar.* So fix it and stop going in for the low-effort, this-centric clapter.

* Not claiming here that this is unique; there are allowed to be other things that work as the same sort of indicator.

Figuring out the structure of the sentence or code

Since we are not technically coding, which website would we use to find the code in order to analyze the data for us?

How do we enable this while preventing people from accidentally nuking their systems?

This is interesting context. I wonder if that need has gone away with large screens or if we're not using it the way it was originally intended. My intuition is that auto-layout is generally better but for smaller pieces of data ad hoc overlaps seem fine.

Reviewer #3 (Public Review):

Truman et al. investigated the contribution and remodeling of individual larval neurons that provide input and output to the Drosophila mushroom body through metamorphosis. Hereto, they used a collection of split-GAL4 lines targeting specific larval mushroom body input and output neurons, in combination with a conditional flip-switch and imaging, to follow the fates of these cells.

Interestingly, most of these larval neurons survive metamorphosis and persist in the adult brain and only a small percentage of neurons die. The authors also elegantly show that a substantial number of neurons actually trans-differentiate and exert a different role in the larval brain, compared to their final adult functionality (similar to their role in hemimetabolous insects). This process is relatively understudied in neuroscience and of great interest.

Using the ventral nerve cord as a proxy, the authors claim that the larval state of the neuron would be their derived state, while their adult identity is ancestral. While the authors did not show this directly for the mushroom body neurons under study, it is a very compelling hypothesis. However, writing the manuscript from this perspective and not from the perspective of the neuron (which first goes through a larval state, metamorphosis, and finally adult state), results in confusing language and I would suggest the authors adjust the manuscript to the 'lifeline' of the neuron.

In general, this manuscript does not explain how the larval brain has evolved as the title suggests but instead describes how the larval brain is remodeled during metamorphosis. It thus generates perspectives on the evolution of metamorphosis, rather than the larval state. Additionally, this manuscript would benefit from major rearrangements in both text and figures for the story to be better comprehended.

The introduction is very focused on the temporal patterning of the insect nervous system, while none of the data collected incorporate this temporal code. Temporal patterning comes back in the discussion but is purely speculative.

Furthermore, the second part of the introduction describes one strategy for remodeling and why that strategy is not likely but does not present an alternative hypothesis. The first section of the results might serve as a better introduction to the paper instead, as it places the results of the paper better and concludes with the main findings. The accompanying Figure 1 would also benefit from a schematic overview of the larval and adult mushroom bodies as presented in Fig. 2A (left).

In the second results section, the authors show the post-metamorphic fates of mushroom body input and output neurons and introduce the concept of trans-differentiation. Readers might benefit from a short explanation of this process. I also encourage the authors to revisit this part of the text since it gives the impression that the neurons themselves undergo active migration (instead of axon remodeling).

The discussion starts with a very comprehensive overview of the different strategies that neurons could use during metamorphosis (here too, re-writing the text from the neurons' perspective would increase the reflection of what actually happens to them).

The discussion covers multiple topics concerning trans-differentiation, metamorphosis, memory, and evolution and is often disconnected from the results. It could be significantly shortened to discuss the results of the paper and place them in current literature. Generally, the figures supporting the discussion are hard to comprehend and often do not reflect what the text is saying they are showing.

Atari, M., Reimer, N. K., Graham, J., Hoover, J., Kennedy, B., Davani, A. M., Karimi-Malekabadi, F., Birjandi, S., & Dehghani, M. (2021). Pathogens Are Linked to Human Moral Systems Across Time and Space. PsyArXiv. https://doi.org/10.31234/osf.io/tnyh9

!- installation : example apps | Peergos - From the top of this page click the green 'Code' button and select 'Download Zip'. - Unzip file and upload desired application folder to Peergos - Navigate into application folder, open the context menu for the file 'peergos-app.json' and choose 'Install App' - Make sure to take note of the file associations and permissions requested - Installed Apps are displayed on the Launcher page (top icon in left menu)

Not easy to read this, it looks like some function f().exp(a)

Replace last sentence with "Use exp(a) to compute", to separate code and full stop.

I think its important to acknowledge that NFTs aren't ever actually owned by players, they are owned by smart contracts and subject to any code in that smart contract. Players have the right to sign transactions in that smart contract involving NFTs marked as being assigned to their wallet address but its really only an illusion of ownership!

important de spécifier cet attribut, afin que javascript ne tente pas d'exécuter ce code, contenant du JSX qui ne sera pas "compris" par JavaScript...

Ce script sera au contraire exécuté par la librairie Babel, chargée plus haut

JSX est un moyen d'écrire du HTML directement dans notre code JavaScript

-> nous verrons plus en détail comment cela fonctionne...

sourcegraph

@palu : Let's highlight that only 1 transfer request can be created in one API call

@palu: Same comments as I made in the transfer entity comments

• don't do p2p stuff directly in the browser
• for privacy reason
• do not broadcast ip address

Martinique criminalizes homosexuality which clashes with France's laws

• [ ] 值得学习git 回滚

Use a lot in code. They are called "if statements"

• SEE

[[Video: Joe Armstrong - Keynote: The Forgotten Ideas in Computer Science - Code BEAM SF 2018]] [[@Joe Armstrong]]

numbering breaking after the code snippet insert. Please check and fix the formatting.

This solution (https://stackoverflow.com/a/47448257/2179543) works fine. My experience on Hubble2NN after having chipset and keywords screening the found TRs only 2 or 3 and then somehow the df to corpus convertion got an error because text columns should be in meta section but went to feature incorrectly. This sample code works immediately to have fixed the problem.

It seems like there might be reason to shift from one code architecture to another, or some series of software design principles that are idiomatic in the target language but not in the source language, for fear of creating code that's not maintainable or easily understood by those familiar with work in the latter programming language. Does this matter if we're going to hide most of the work behind abstraction boundaries though?

How?

How do we validate that the code does the same thing? I'm not sure that manual review is sufficient - there are too many "gotchas" here. How can we bring in contracts, tests or other semi-verified methods of confirming that computation is accurate, operating in a (relatively) language agnostic way? Is this even possible?

Author Response

Reviewer #1 (Public Review):

In their manuscript, these authors present a novel geostatistical framework for modelling the complex animal-environment-human interaction underlying Leptospira infections in a marginalised urban setting in Salvador, Brazil.

In their work, the authors combine human infection data and the rattiness framework of Eyre et al. (Journal of the Royal Society Interface, 2020) . They use seroconversion defined as an MAT titer increase from negative to over 1:50 or a four-fold increase in titer for either serovar between paired samples from cohort subjects. Whereas this is a commonly used measure of infection; the work would benefit from answering the question about how robust results are related to this definition of seroconversion.

Thank you for your comment. We have acknowledged this on line 534 in the discussion by adding the following text: “A possible limitation of this study is the titre rise cut-off values used for classifying seroconversion and reinfection in the cohort that determine the sensitivity and specificity of the infection criteria. However, these criteria were used because they are the standard definitions for serological determination of infection that are commonly applied for leptospirosis and a wide range of other infections, and they enable the comparison of results with other previous leptospirosis studies.”

The model framework relies on the concept of 'rattiness' previously defined by Eyre et al. (JRSI, 2020) and assumes conditional independence within its built up (equation (1)). Whereas this is a reasonable assumption, it would be good to discuss situations in which this assumption is questionable and what the implications are for applying the modelling framework to other settings.

We have added the following text immediately after “is shown schematically in Figure 2” following equation (1) on line 225: “The conditional independence assumption in (1) is reasonable for a vector-borne disease or one that is transmitted indirectly, in which context the observed rat indices are to be considered as noisy indicators of the unobservable spatial variation in the extent to which the environment is contaminated with rat-derived pathogen. It would be more questionable for applications in which the disease of interest is spread by direct transmission from rat to human.”

The authors provide an extensive model building exercise and investigate, in different ways, whether the model captures the necessary complexity (GAM smoothers - testing linearity, spatial correlation, etc). I believe the work would benefit from (1) a formal diagnostic investigation, if feasible; (2) providing guidelines on how model building should be performed.

We have added a new Appendix 7 with diagnostic plots of randomized quantile residuals to check the rattiness-infection model fit with the human infection data and included the following text in Section 2.4 of the main text: “A formal diagnostic investigation of randomized quantile residuals is included in Appendix 7. We found no evidence in the diagnostic plots to suggest that there were issues with our modelling approach.”

To supplement the R code that is publicly available for repeating all of the steps in this analysis, we have now also included a detailed step-by-step explanation of the model building process in Appendix 8 that outlines the key steps for building the rat and infection components of the model (variable selection and evaluation of residual spatial autocorrelation) and fitting and examining the joint rattiness-infection model. We have added the following text in Section 2.6 of the main text: “We also include a step-by-step explanation of the model building process to guide future users of the rattiness-infection framework in Appendix 8.”

The authors are to be acknowledged for providing an extensive and thorough discussion of the different aspects of their work. Whereas the discussion is complete, I wonder whether the authors can give a brief example about how this model can be applied in a different setting.

Thank you. We have added the following text on line 551 in the discussion: “The framework may have important applications beyond the study of zoonotic spillover, with the rattiness component replaced by other exposure measures e.g. mosquito density or ecological indices (such as pollution, where there are multiple, related measures of air or groundwater quality) to model associations with human or animal health outcomes.”

Reviewer #2 (Public Review):

Eyre et al. developed and applied a novel geostatistical framework for joint spatial modeling of multiple indices of pathogen (Leptospira) reservoir (rats) abundance and human infection risk. This framework enabled evaluation of infection risk at a fine spatial scale and accounted for uncertainty in the pathogen reservoir abundance estimates. The authors used data collected in two different field projects: (1) a rat ecology study in which three different approaches were used to detect rat presence "rattiness", and (2) a prospective community cohort study in which individuals were sampled during two different time periods to detect recent infections via seroconversion or a four-fold increase in anti-Leptospira antibody MAT titer. Univariable and then multivariable analyses were performed on these data to identify (1) the environmental variables that best predicted "rattiness", and (2) the demographic/social, environmental (household), occupational, and behavioral variables that best predicted human risk of infection. Once identified, the best predictors from (1) and (2) were included in a final, joint model to identify the significant predictors of both 'rattiness' and human infection risk. As a result of this study, the authors were able to detect spatial heterogeneity in leptospiral transmission to humans. They found that infection risk associated with increases in reservoir abundance differed by elevation, and that increases in reservoir abundance at high elevation were associated with a much higher odds ratio for infection than at low elevation. The authors suggest that this has to do with differences in how the infectious leptospires (shed by the rat reservoir) are dispersed in the environment. At high elevations, flooding is less frequent and thus rat shed leptospires are likely to stay where the rat deposited them. Whereas at lower elevations, flooding may play a large role in spreading leptospires more evenly across the landscape, reducing the importance of rat presence at smaller spatial scales. The final best model was then used to generate prediction maps of 'rattiness' as well as human infection risk at all locations within the study area (i.e. including those that lacked rat detection data and human infection data. This work represents an important advance in infection risk modeling as it explicitly incorporates estimates of reservoir abundance and the uncertainty surrounding these estimates into the infection risk assessment, and allows for modeling of infection risk at fine spatial scales. Findings from this study have important management implications at the authors' study site as it suggests that interventions directed at high elevations should be different from those designed to address infection risk at lower elevations. However these are broader implications, as this novel approach may be applied to other systems to enable identification of differences in infection risk for other pathogens at a fine spatial scale, predict infection risk more broadly, and facilitate intervention strategies targeted for the specific epidemiological and ecological conditions experienced by a population.

This was a well-designed study. The field sampling approach was well balanced, well described and appropriate. Broadly the modeling framework is appropriate for the questions being asked and for the data being used. The variable and model selection approaches were clearly described and appropriate. Evaluation of the more detailed mathematical approach is outside of my area of expertise, so I am unable to comment on the validity of the approach.

For the most part, the explanatory variables assessed in the different models were well described and justified, however there were some cases for which further explanation would have been helpful. For example, how did the authors determine which occupations to evaluate? Specifically, why traveling salesperson? What is the difference between open sewer within 10 m and unprotected from sewer?

We have added the following additional text to Section 2.3.2 on line 297 to clarify the definition and reason for inclusion for these variables: “In the household environment domain, two variables were used to capture risk due to sewer flooding close to the household: i) the presence of an open sewer within 10 metres of the household location and ii) a binary `unprotected from open sewer' variable which identified those households within 10 metres of an open sewer that did not have any physical barriers erected to prevent water overflow. Three high-risk occupations were included in the occupational exposures domain as binary variables. Construction workers and refuse collectors have direct contact with potentially contaminated soil, building materials and refuse in areas that provide harbourage and food for rats. Travelling salespeople have regular and high levels of exposure to the environment (particularly during flooding events) as they move from house to house by foot. Two other binary occupational exposure variables were included that measured whether a participant worked in an occupation that involves contact with floodwater or sewer water.”

I also had some concerns regarding the time-period of the rat ecology study used to determine abundance, potential fluctuations in rat abundance through time, and how this might align with sampling to detect infection in humans. Depending on the time scale of population fluctuation in rats as well as fluctuations in infection prevalence in rats, the abundances calculated from data from the ecology study may not be accurately reflecting true abundance (and therefore shedding and transmission risk) during the time period that a human may have been exposed. However, the authors do a nice job of addressing some of these issues in the discussion. They mention that infection prevalence in rats is consistently around 80% and that there don't appear to be seasonal fluctuations in human exposure risk in the study area.

Thank you.

Reviewer #3 (Public Review):

The goal of the authors was to test how important local rat abundance is as a driver of Leptospira infection in humans.

The authors approached this using a strong combination of datasets on human infection risk and rat abundance, across a spatial scale that is large enough to allow simultaneous assessment of multiple potentially important drivers of infection risk. This further enables the authors to develop infection prediction maps based on the fitted models.

This study design is a major advance towards understanding link between rat abundance and human infection risk.

Based on the top models tested in the study, the authors conclude that local rat abundance is indeed correlated with infection risk, and that this correlation is strongest at higher elevation.

This is an impactful finding, but in my opinion it is not yet clear how robust and important this is, because of two reasons:

(1) The infection risk data: while the actual infection risk data are not shown, the map shown in Figure 5B suggests that there is an infection hotspot that happens to be at high elevation. This raises the question of how strongly this single hotspot is driving the observed correlation between rat abundance and infection risk (which the authors find to be much stronger at high elevation than at lower elevations).

We have added a new figure (Figure 4) earlier on in the article (we decided to add this here rather than to Figure 6 - formerly Figure 5 - to ensure that the map is large enough that points in Figure 4A are easily visible – please note that it is included as a larger and easier to view image in the main eLife template version) with the raw infection data overlaid on contour lines for the three elevation levels to provide the reader with a better overview of the raw data. This new Figure 4 shows that out of a total of 403 participants in the high elevation region there were 16 infections, of which only 5 (31%) were located in the large hotspot in Valley 3 (valleys are numbered 1 to 3 from west to east, see Figure 1A). In addition to the largest hotspot in the north of Valley 3, there are several other areas in the high elevation region with raised predicted infection risk values relative to their surroundings where there were also rattiness hotspots and infected participants in the raw data: fives cases (red and yellow infection risk areas in Figure 5B) on the western side of Valley 2; the two cases on the eastern edge of Valley 2; the two cases on the western edge of Valley 3; and the single case in the southwest of Valley 3. Other variables are also important drivers of infection risk and at several of these locations the contribution of rattiness increases infection risk significantly relative to the low-risk surrounding area (e.g. to 10% in areas where risk is closer to 1% or 2%) without reaching the more obviously visible high infection risk values closer to 20%. We believe that our statistical model provides a better test of whether there is a statistical association between rattiness and infection at high elevations than a visual examination, but that this is supported by the large number of observations in the high elevation area (403) and the distribution of infected and uninfected households, which demonstrates that the observed association is not only driven by the hotspot in Valley 2.

(2) The statistical models: if I understand correctly, all tested models of infection risk include the variable rat abundance, and while the individual effect estimates for rat abundance are statistically significant (Table 3), the more important question of how the fit of a model without the rat abundance variables compares with those of the other tested models (shown in Supplementary Table S2) has not been addressed.

These models were considered but were ranked outside of the top five models and for this reason were not reported in Table S2. We agree that showing the AIC of a model without rattiness in this table can more clearly demonstrate the improved fit of the model with rattiness. To do this we have added the highest ranked model without rattiness (M) to Table S2 and added a note to the table explaining the reason for its inclusion (“Model M was ranked outside of the top 5 models but is included here for reference to demonstrate the improvement in model fit when rattiness is included”). The AIC of M* was 532.13. This is substantially higher than the top five models (M1 = 523.14 and M5 = 525.04), justifying its inclusion in this model and in the joint rattiness-infection framework.

Regardless of whether rat abundance is an important driver of human infection risk, this study is a major step in our understanding of the role of rats in the spread of leptospirosis, due to the strong combination of a unique combination of datasets and a spatial statistical modeling approach.

Thank you.

So run a typechecker on the code to check your work if you want type checking. That is what TypeScript does, after all. And it's been around long enough that people shouldn't be making the mistake that a runtime that support dynamic types at runtime means that you can't use a static typechecker at "compile time" (i.e. while the code is still on the developer workstation).

With this comment, the anti-JS position is becoming increasingly untenable. The author earlier suggested C as an alternative. So their contention is that it's easier to write bug-free code in C than it is in JS. This is silly.

C hackers like Fabrice Bellard don't choose C for the things they do because it's easier to write bug-free code in C.

What is the relationship between "arcane code" and "code that you'd have to recompile"?

Where does it say they're "going with web-like scripts"? (And what is "more traditional code"?)

Add the one-liner for Entity

code samples in multiple languages

code samples

code samples?

code samles

code samples

multiple language code

Add code samples in multiple languages

Sample code in multiple languages?

遗留代码

You could make each of these branches its own function. This would shorten the code a bit and make it clearer because you could name the function something informative like: print_help()

Often people will put the stuff at the top level (like this variable declaration, for instance) into a function, then only have one top level statement that looks like:

if name == 'main': my_top_level_function()

Doing so makes it so that other users can import your code without having it run. If you do it like this, the my_top_level_function will only run if it is directly invoked by the interpreter.

https://ncase.me/nutshell/

For those interested in the ideas of annotations, inline footnotes, expandable text, additional context, stretch text, hovercards, etc., Nicky Case has recently released a new project called Nutshell (https://ncase.me/nutshell/) which has some fun user interface as well as code with which to play.

Reviewer #1 (Public Review):

This work describes a new method, Proteinfer, which uses dilated neural networks to predict protein function, using EC terms and GO terms. The software is fast and the server-side performance is fast and reliable. The method is very clearly described. However, it is hard to judge the accuracy of this method based on the current manuscript, and some more work is needed to do so.

I would like to address the following statement by the authors: (p3, left column): "We focus on Swiss Prot to ensure that our models learn from human-curated labels, rather than labels generated by electronic annotation".

There is a subtle but important point to be made here: while SwissProt (SP) entries are human-curated, they might still have their function annotated ("labeled") electronically only. The SP entry comprises the sequence, source organism, paper(s) (if any), annotations, cross-references, etc. A validated entry does not mean that the annotation was necessarily validated manually: but rather that there is a paper backing the veracity of the sequence itself, and that it is not an automatic generation from a genome project.<br /> Example: 009L_FRG3G is a reviewed entry, and has four function annotations, all generated by BLAST, with an IEA (inferred by electronic annotation) evidence code. Most GO annotations in SwissProt are generated that way: a reviewed Swissprot entry, unlike what the authors imply, does not guarantee that the function annotation was made by non-electronic means. If the authors would like to use non-electronic annotations for functional labels, they should use those that are annotated with the GO experimental evidence codes (or, at the very least, not exclusively annotated with IEA). Therefore, most of the annotations in the authors' gold standard protein annotations are simply generated by BLAST and not reviewed by a person. Essentially the authors are comparing predictions with predictions, or at least not taking care not to do so. This is an important point that the authors need to address since there is no apparent gold standard they are using.

The above statement is relevant to GO. But since EC is mapped 1:1 to GO molecular function ontology (as a subset, there are many terms in GO MFO that are not enzymes of course), the authors can easily apply this to EC-based entries as well.

This may explain why, in Figure S8(b), BLAST retains such a high and even plateau of the precision-recall curve: BLAST hits are used throughout as gold-standard, and therefore BLAST performs so well. This is in contrast, say to CAFA assessments which use as a gold standard only those proteins which have experimental GO evidence codes, and therefore BLAST performs much poorer upon assessment.

Pooling GO DAGs together: It is unclear how the authors generate performance data over GO as a whole. GO is really 3 disjoint DAGs (molecular function ontology or MFO, Biological Process or BPO, Cellular component or CCO). Any assessment of performance should be over each DAG separately, to make biological sense. Pooling together the three GO DAGs which describe completely different aspects of the function is not informative. Interestingly enough, in the browser applications, the GO DAG results are distinctly separated into the respective DAGs.

Figure 3 and lack of baseline methods: the text refers to Figures 3A and 3B, but I could only see one figure with no panels. Is there an error here? It is not possible at this point to talk about the results in this figure as described. It looks like Figure 3A is missing, with Fmax scores. In any case, Figure 3(b?) has precision-recall curves showing the performance of predictions is the highest on Isomerases and lowest in hydrolases. It is hard to tell the Fmax values, but they seem reasonably high. However, there is no comparison with a baseline method such as BLAST or Naive, and those should be inserted. It is important to compare Proteinfer with these baseline methods to answer the following questions: (1) Does Proteinfer perform better than the go-to method of choice for most biologists? (2) does it perform better than what is expected given the frequency of these terms in the dataset? For an explanation of the Naive method which answers the latter question, see: (https://www.nature.com/articles/nmeth.2340)

Reviewer #2 (Public Review):

In this paper, Sanderson et al. describe a convolutional neural network that predicts protein domains directly from amino acid sequences. They train this model with manually curated sequences from the Swiss-Prot database to predict Enzyme Commission (EC) numbers and Gene Ontology (GO) terms. This paper builds on previous work by this group, where they trained a separate neural network to recognize each known protein domain. Here, they train one convolutional neural network to identify enzymatic functions or GO terms. They discuss how this change can deal with protein domains that frequently co-occur and more efficiently handle proteins of different lengths. The tool, ProteInfer, adds a useful new tool for computational analysis of proteins that complements existing methods like BLAST and Pfam.

The authors make three claims:<br /> 1) "ProteInfer models reproduce curator decisions for a variety of functional properties across sequences distant from the training data"<br /> .<br /> This claim is well supported by the data presented in the paper. The authors compare the precision-recall curves of four model variations. The authors focus their training on the maximum F1 statistic of the precision-recall curve. Using precision-recall curves is appropriate for this kind of problem.

2) "Attribution analysis shows that the predictions are driven by relevant regions of each protein sequence".

This claim is very well supported by the data and particularly well illustrated by Figure 4. The examples on the interactive website are also very nice. This section is a substantial innovation of this method. It shows the value of scanning for multiple functions at the same time and the value of being able to scan proteins of any length.

3) "ProteInfer models create a generalised mapping between sequence space and the space of protein functions, which is useful for tasks other than those for which the models were trained."

This claim is also well supported. The print version of the figure is really clear, and the interactive version is even better. It is a clever use of UMAP representations to look at the abstract last layer of the network. It was very nice how each sub-functional class clustered.

The interactive website was very easy to use with a good user interface. I expect will be accessible to experimental and computational biologists.

The manuscript has many strengths. The main text is clearly written, with high-level descriptions of the modeling. I initially printed and read the static PDF version of the paper. The interactive form is much more fun to read because of the ability to analyze my favorite proteins and zoom in on their figures (e.g. Figure 8). The new Figure 1 motivates the work nicely. The website has an excellent interactive graphic showing how the number of layers in the network and the kernel size change how data is pooled across residues. I will use this tool in my teaching.

The manuscript has only minor weaknesses. It was not clear if the interactive model on the website was the Single CNN model or the Ensemble CNN model.

Overall, ProteInfer will be a very useful resource for a broad user base. The analysis of the 171 new proteins in Figure 7 was particularly compelling and serves as a great example of the utility and power of ProteInfer. It completes leading tools in a very valuable way. I anticipate adding it to my standard analysis workflows. The data and code are publicly available.

Reviewer #2 (Public Review):

In this manuscript, Ibanez-Sole et al. focus on an important open question in ageing research; "how does transcriptional noise increase at the cellular level?". They developed two python toolkits, one for comparison of previously described methods to measure transcriptional noise, Decibel, and another one implementing a new method of variability measure based on cluster memberships, Scallop. Using published datasets and comparing multiple methods, they suggest that increased transcriptional noise is not a fundamental property of ageing, but instead, previous reports might have been driven by age-related changes in cell type compositions.

I would like to congratulate the authors on openly providing all code and data associated with the manuscript. The authors did not restrict their paper to one dataset or one approach but instead provided a comprehensive analysis of diverse biology across murine and human tissues.

While the results support their main conclusions, the lack of robustness/sensitivity measures for the methods used makes it difficult to judge the biology. The authors use real data to compare between methods but using synthetic data with known artificial 'variability' across cell clusters can first establish the methods, which would make the results more convincing and easier to interpret. Despite the comprehensive analysis of biological data, a detailed prior description of how the methods behave against e.g. the number of cells in each cell type cluster, the number of cell types in the dataset, and % feature expression, would make the paper more convincing. Once the details of the method is provided, the python toolkit can be widely used, not limited to the ageing research community. I am also concerned that a definition of 'transcriptional noise' (e.g. genome-wide noise, transcriptional dysregulation in cell-type-specific genes, noise in certain pathways) and its interpretation with regard to the biology of ageing is missing. Differences in different methods could be explained by the different biology they capture. Moreover, the interpretation of a lack of different types of variability may not be the same for the biology of ageing.

Increased transcriptional noise is compatible with genomic instability, loss of proteostasis and epigenetic regulation. Showing a lack of consistent transcriptional noise can challenge the widespread assumptions about how these hallmarks affect the organism. Overall, I found the paper very interesting and central to the field of ageing biology. However, I believe it requires a more detailed description of the methods and interpretations in the context of biology and theories of ageing.

Major comments:<br /> 1. The concept of transcriptional noise is central to the manuscript; however, what the authors consider as transcriptional noise and why is not clear. Genome-wide vs. function or cell-type specific noise could have different implications for the biology of ageing. In line with this, a discussion of the findings in the context of theories of ageing is necessary to understand its implications.<br /> 2. While I found the suggested method, Scallop, quite exciting and valuable, I would suggest including a number of performance/robustness measures (primarily based on simulations) on how sensitive the method is to the number of cells in each cell type (cellular composition), misannotations, % feature expression (number of 0s) etc.:<br /> 2.1. Most importantly, knowing that cell-type composition changes with age, it is important to know how sensitive community detection is to the number of cells in each cell type. While the average can be robust, I wonder if the size of the cell-type cluster affects membership (voting).<br /> 3. Although the Leiden algorithm is widely used by many single-cell clustering methods, since the proposed methodology is heavily dependent on clustering, I suggest including a description of the Leiden algorithm.<br /> 3.1. Most importantly, the authors comment that they found stronger expression of cell-type specific markers in the cells with high membership values - is it already a product of the Leiden algorithm that it weighs highly variable (thus cell-type specific) features higher - resulting in better prediction of cell-types for cells with strong cell-marker expression? It is important to make a description of transcriptional noise at this stage as it could be genome-wide or more specific to cell-type markers. Can authors provide any support that their method can capture both?<br /> 4. The authors conclude that Scallop outperforms other methods through the analysis of biological data, where there is no positive and negative control. I suggest creating synthetic datasets (which could be based on real data), introducing different levels of noise artificially (considering biological constraints like max/min expression levels) and then testing the performance where the truth about each dataset is known. Otherwise, the definitions of noisy and stable cells, regardless of the method, are arbitrary.

Was ich vermisse: "rechtliche Rahmenbedingungen"

Heißt das nun endlich auch, dass eine Förderung durch das BMBF an Offenheit als Voraussetzung geknüpft wird? Zum Beispiel "public money, public code", Material verpflichtend als OER, Publikationen unter offener Lizenz, Open Science, ...

Reviewer #3 (Public Review):

Numerous studies have demonstrated that the neural dynamics on different brain areas encode elapsed time, yet it has proven challenging to examine how these population clocks emerge over the course of learning because most temporal tasks require many training sessions. In this manuscript the authors use a simple timing task that can be learned in a single day, and accompany the changes in neural dynamics in the mPFC and STR of the first and second day on the task. The most interesting finding is a switch in which the mPFC provides a better code than the STR for elapsed time on the first day, but the STR provides a better code than the mPFC on the second day. Consistent with the increased encoding of time in the mPFC early in training, muscimol inactivation of the mPFC impaired learning of the task, but not performance in trained animals. Overall this study provides a number of novel contributions to our understanding of temporal processing, and show the first example of learning-dependent switch from the dynamics of the mPFC to that of the STR encode time.

Evaluation Summary

This study investigates the question of whether distinct brain areas differentially encode time during the learning of a simple motor timing task. The key novel result is that early in training the dynamics of the medial prefrontal cortex (mPFC) provides the best code for time, but later in training, the basal ganglia and in particular, the striatum provides a better code. In addition, the study shows that inactivation of mPFC produces a delayed learning effect, while inactivation of the striatum after learning led to impaired performance. The observation that temporal coding and the necessity of brain area for task performance transfers from medial prefrontal cortex to the striatum during learning is an intriguing observation for our understanding of the neural mechanisms underlying temporal processing in sensorimotor control.

(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

This note (and most modern JS programmers' code) could benefit from the eradication of its aversion to this.

I agree with this. I hate the config code and I get away with using almost none of it.

You can do this in 11ty without configuration code with folder structures and directory data files.

Does it make sense to class the big'uns (Jekyll, Hugo, etc.) with somebody's bash script as the same kind of thing? Is there really more use of the latter?

Reviewer #3 (Public Review):

Understanding how neural representations throughout the brain, including the hippocampus, interact with neuromodulators such as acetylcholine to support flexible and lasting episodic memories is a fundamental question of interest to a broad neuroscientific community. Here, Blair et al. build on existing literature to concurrently characterize the relationships among these elements. Using large-population widefield miniscope recordings combined with systemic scopolamine administration in rats, the authors first demonstrate that localized aversive experiences result in lasting avoidance behavior as well as changes to (a.k.a. 'partial remapping of') the hippocampal neural code, with lasting changes occurring predominantly near the aversive experience, all replicating prior work with high precision. Next, the authors show that systemic administration of the acetylcholine antagonist scopolamine during the aversive experience gives rise to a different but reliable hippocampal code during that experience. Moreover, rats on scopolamine did not exhibit lasting avoidance behavior or changes to their hippocampal codes from before or after the experience, suggesting that the instantiation of a different hippocampal code during the aversive experience shielded the existing representation and its associated behavior from experience-induced changes. Together, these results demonstrate novel, provocative links between episodic memory, the plasticity of hippocampal neural codes, and the neuromodulator acetylcholine, with a number of important implications for how this memory system functions.

In my eyes, this work has a number of strengths. One major strength is the power and precision afforded by the use of the large-field miniscope recordings. While this may leave questions of fine temporal structure unaddressable, many of the questions of interest here are best addressed with large populations of simultaneously-recorded neurons that can be confidently tracked across at least a week, all of which are strengths of this technique. Another strength of this work is the replicate and extend approach to addressing the relationships among this work's components. The links to prior work in all of these cases are well noted, the replications of prior results are often with significantly more statistical power than the original result had, and these replications raise confidence in the quality of the data and the novel results reported here.

As with all work, this too has its limitations. One fundamental limitation is the inability to speak to functional localization. That is, although this work points to provocative correlational links among acetylcholine, the plasticity of hippocampal codes, and behavioral memory expression (all of which are well-motivated by existing literature) because the administration of scopolamine is systemic and only one region can be monitored it is impossible to draw causal conclusions from this work. While it is tempting to infer that manipulating acetylcholine modulation of hippocampal plasticity is necessary and sufficient to produce these results, it is also possible that the behavioral impact of the acetylcholine manipulation is driven by regions outside of the hippocampus and that changes to the hippocampal plasticity are not behaviorally relevant, or that these changes are necessary but not sufficient to drive memory expression. A specific version of this limitation is referenced by the authors in the discussion when considering the possible impact of the manipulation on amygdala responses.

Despite its limitations, this work meaningfully complements and extends existing literature probing the links between episodic memory, the plasticity and stability of hippocampal codes, and neuromodulators such as acetylcholine.

issue: whether she qualifies for the immunity that a social host gets or whether she gave it up by selling to a minor"

Clean code?

X for thee, not for me

Applicable as criticism for: - marketing analytics for "data-driven" campaigns - the sort of resume-driven development that is fashionable on Github - naive code reuse, generally

This is still not the best way to put it. Decisions come at the expense of writing code. This maintenance work is something else.

This is just such a great title!

Evaluation Summary:

This study will interest basic and clinical scientists and, potentially, device manufacturers interested in the regulation of heart rate in health and disease. A major control of the heart is from the nervous system originating in a neuronal cluster sitting outside the heart called the stellate ganglia. This study has identified the neural code associated with a healthy heart and describes how it changes in disease. Whether the change in code is cause or effect remains equivocal although normalising the code may have valued therapeutic benefit. The study opens the way for sophisticated mimicking of healthy neural code applied to a diseased heart as a potential electroceutical approach.

(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)

Reviewer #1 (Public Review):

The authors have performed a deep mathematical analysis of unitary data recorded from the stellate ganglion to understand how the neural code is altered in heart failure.

The study is advantaged by being performed in vivo with afferent and efferent pathways intact. The use of modern microelectrode arrays has allowed mass activity to be recorded from multiple sites simultaneously within the ganglion. The authors have a number of powerful analytical tools that have revealed quantitative changes of interest.

The data are from animals under anesthesia with an open chest and open pericardial sac and one wonders what effect this has on the neural activity given the changes in pulmonary physiology this will cause.

Some of the data are from pigs where resiniferatoxin (a chemical agent to kill sensory afferents) was applied to the epicardial surface. Given the elevation in sensitivity of cardiac afferent reflexes in heart failure (Schultz, Zucker, and others), it is surprising that this had no effect on the neural activity recorded in the heart failure animals. Either the afferents were not destroyed (no data given to demonstrate this) or these sensory fibres play no role in the changes in neural activity reported from heart failure pigs. This would go against current data and remains unclear.

Most of the stellate neurons project to non-cardiac tissues. One does not get a sense of the proportion of activity that was related to the heart (left ventricular pressure) and whether in heart failure there is an elevated activity within a confined network or recruitment of additional networks. In this regard, the manuscript is jargon-heavy and for those that are physiologists, the subtleties of the study may be lost.

Finally, the authors could provide a clearer take-home message and break out of the shackles of math talk and interpret the possible physiological relevance of the work as well as why it is important to understand the changes in stellate neural network dynamics in heart failure.

Author Response

Reviewer #1 (Public Review):

In this paper, Jan Kubanek attempts to derive an 'effective decision strategy' that is optimal (and therefore normative) given certain constraints resulting from computational capacity limitations. The author first points out that neoclassical economics (i.e., expected utility theory, EUT) provides normative predictions for decisions to maximize utility. Next, he (correctly) points out that finding the optimal solutions to decision problems requires computational resources that are unlikely to exist in actually existing decision-makers (animals and humans). He claims that this fact is the most severe problem for concluding that EUT is an accurate description of actual human or animal decision processes. I disagree with him on this point as I will lay out in more detail below. Next, the author attempts to find an 'efficient' (i.e., computationally reasonable) decision strategy that comes close to the original normative framework. He claims that such a strategy is EDM, whereby decisions are made by allocating relative effort in proportion to the relative reward of each option.

Overall, I find this paper hard to judge. The considerations described in this paper are certainly interesting and I have no reason to presume that the mathematical derivations described are wrong (without having made an effort to follow and check it in detail). Still, I find the paper, in the end, sterile and I fear it will have only limited impact. I think the manuscript should be expanded in three different directions to make it more relevant for the neuroscientific understanding of decision making.

First, the author needs to show that EDM can also explain other known violations of EUT related to the axiom of regularity (i.e., preferences between two options should not be affected by the presence of inferior options). This seems relevant because these behavioral effects robustly violate the choice allocation strategy of EDM.

Second, EDM is so abstract that the actual structure and capacity of the nervous system are nearly irrelevant. The author should consider more deeply the computational requirements and capacities of different types of brains; fruit flies, frogs, and primates, and the consequences of these differences for what is (or should be) achievable in terms of optimal behavior.

Third, the paper contains no test for EDM. This is in part because EDM is at no point compared to the predictions of alternative theories.

I thank the Reviewer for these constructive comments, which are addressed below.

My specific concerns are as follows:

(1) The author claims that the most severe problem of EUT is that it is computationally implausible. However, I disagree.

It could be claimed that EUT describes an (unattainable) optimal state that actual brains try to accomplish with limited resources. (In essence, the current paper follows this strategy).

Correct, EDM stems from Expected Utility Theory subjected to specific biological considerations, as shown in Figure 1.

Given this origin, the paper now makes more appropriate statements regarding the biologically-relevant shortcomings of EUT:

i) Abstract: "the apparatus requires a large number of evaluations of the decision options as well as neural representations and computations that are not biologically plausible."—>"the apparatus requires a large number of evaluations of the decision options as well as neural representations and computations that are difficult to implement at the biological level" ii) Introduction: "To address these biologically implausible requirements, …" —> "To address the biological constraints, …").

I think the situation is much direr. During the last 70 years, a small army of psychologists and behavioral economists have described a large number of violations of EUT's normative predictions: the Allais paradox, framing effects, the behavioral tendencies summarized in Prospect theory, and others. These differences between behavior and normative predictions are important because they violate basic assumptions of the normative theory.

Prospect theory can be readily incorporated into EDM.

This has resulted in the following paragraph in the Discussion:

"Notably, the 𝑢𝑖 and 𝑒𝑖 variables can incorporate additional factors such as the probability of an outcome, as in prospect theory (Kahneman and Tversky, 1979). A previous study (Kubanek, 2017) demonstrates that prospect theory’s incorporation of probabilities into utilities does not change the relationship between the differential formulation of Equation 1 and the fractional formulation of Equation 2, which is crucial for EDM. Moreover, the 𝑢𝑖 and 𝑒𝑖 variables can be entirely subjective. So long as the representations are comparable by the brain (e.g., through relative firing rates; Figure 6), the 𝑒𝑖 = 𝑢𝑖 strategy provides an efficient allocation of the decision-maker’s resources."

(2) The most interesting case of such violations is a set of well-known behavioral effects that occur in the context of multi alternative-multi attribute decision making. They are known as the attraction, similarity, and compromise effects (there is a large literature; more recently: Dumbalska T, Li V, Tsetsos K, Summerfield C. A map of decoy influence in human multi alternative choice. Proc Natl Acad Sci U S A. 2020 Oct 6;117(40):25169-25178. doi: 10.1073/pnas.2005058117. Epub 2020 Sep 21.) These biases have received so much attention because they violate a very basic axiom of EUT. Choices between two options should not be affected by the presence of a third option that is inferior to both of them. However, that is exactly what happens in these choice biases. The effects have been shown in many species ranging from humans to amphibians to invertebrates. As far as I can see, EDM cannot explain how choice allocation between two options A and B that have equal value would be changed by the inclusion of a new option D so that is of lower value than A or B in such a way that D is not chosen at all, but A is chosen more often than B if D is similar in attributes to A (the 'attraction' effect). If I am mistaken, the inclusion of an explanation of how this would work would be of major importance.

The new Figure 6 provides a starting point for addressing these effects.

Specifically, this comment has resulted in the following Discussion paragraph:

"In EDM, the relativistic representation of utility at the neural level (black bars in Figure 6) involves divisive normalization. Divisive normalization a common operation performed by neural circuits (Carandini and Heeger, 2012). The specific form of this operation may be crucial for explaining attraction, similarity, and compromise effects observed in multi-alternative, multi-attribute decision environments (Noguchi and Stewart, 2014; Dumbalska et al., 2020). For instance, it has been found that a transformation of utilities by specific monotonic functions prior to divisive normalization can explain these behavioral effects parsimoniously (Dumbalska et al., 2020). On this front, monotonic transformations and divisive normalization are performed by several kinds of feedforward and feedback neural circuits (Lek et al., 1996; Carandini and Heeger, 2012). Nonetheless, how exactly individual attributes of decision options are encoded at the neural level should be investigated using large-scale neuronal recordings."

(3) EDM as described in this manuscript is completely static, that is it ignores actual computational processes that underlie decision making. This is in opposition to an important modern branch of decision research that has stressed the importance of understanding processes (and their limitations) to understand how choices are made. Examples are: (1) Roe RM, Busemeyer JR, Townsend JT. Multialternative decision field theory: a dynamic connectionist model of decision making. Psychol Rev. 2001 Apr;108(2):370-92. doi: 10.1037/0033-295x.108.2.370. PMID: 11381834.; (2) Tsetsos K, Usher M, Chater N. Preference reversal in multiattribute choice. Psychol Rev. 2010 Oct;117(4):1275-93. doi: 10.1037/a0020580. PMID: 21038979. The relationship between EDM and algorithmic implementations should be explored.

This point has been addressed in the following ways:

1) EDM is now implemented at the algorithmic level while positioned within stochastic choice environments.

2) The performance of EDM in the stochastic environments is reported in a new Figure 4.

3) The performance of EDM within the stochastic and deterministic environments is now compared in a new Figure 5. The figure shows that both environments support the same principal conclusions.

4) Figure 4b provides mechanistic examples of the individual effort allocations by EDM and alternative strategies.

5) The Discussion includes a new paragraph that places EDM within a broader context of algorithmic implementations:

"In deterministic environments, EDM comprises a single stage that embodies Equation 7. This rule is analogous to the evolutionarily stable “relative reward sum” in ecology (Harley, 1981; Hamblin and Giraldeau, 2009) and “local fractional income” in neuroscience (Sugrue et al., 2004). In dynamic and stochastic environments, the strategy should additionally incorporate an integration stage that mitigates the effect noise and thus provides meaningful estimates of the worth of each option. Several approaches can be used to keep track of dynamic, stochastic environments and thus estimate their relative worth 𝑢𝑖. The most compact are related to reinforcement learning, in which previous payoffs are discounted exponentially using a “learning rate.” This approach has been applied in ecology (Harley, 1981; Hamblin and Giraldeau, 2009), computer science (Sutton and Barto, 1998), neuroscience (Sugrue et al., 2004; Corrado et al., 2005), and was also applied here when assessing performance in stochastic environments. One benefit of this free parameter is that decision-makers can adapt the learning rate to the speed of change or the level of stochasticity of specific decision situations (Iigaya et al., 2019)."

(4) Most importantly, what is missing is a clear prediction for a finding (behavioral or neuronal) that would only be predicted, but not by any other theory of decision making. Without such a proposed test, the idea has no scientific merit.

The paper includes new analyses and text that provide predictions that are specific to EDM. Specifically, this point has been addressed in three ways:

1) The three predictions that are specific to EDM are now made explicit in a new Figure 5. The figure also provides quantitative support of EDM through performance evaluations across these predictions.

2) The Results include the following text regarding the key defining properties of EDM: "Figure 5 summarizes and expands on the defining properties of EDM. First, the main finding of this article is that EDM is characterized by high performance following a single evaluation of decision options (Figure 5a). Second, Figure 3a suggested that the proportional allocation of effort to relative utilities (𝛽 → 1) may represent an optimum, at least across the space of effort-utility contingencies tested. Figure 5b-top additionally evaluates the impact of this exponent in the stochastic choice situations. This figure replicates the findings of Figure 3a in that 𝛽 = 1 lies near the optimum, with 𝛽 = 1.0 and 𝛽 = 1.2 providing an average gain of 94.0% and 94.1%, respectively. Thus, the proportional allocation of effort to relative utilities is another defining trait of EDM, and this strategy provides near-optimal performance in all decision situations tested. And third, the effort allocation strategy in EDM, 𝑒harvest = 𝑢(𝑒eval), is invoked once regardless of the number of decision options. This is in contrast to optimization, whose convergence time scales with problem dimensionality, i.e., the number of options. The single-evaluation EDM strategy maintains performance across the number of options under the VI schedules (Figure 5c-top; slope 0.67% per option, 𝐹 = 4.46, 𝑝 = 0.073), although it does incur a performance loss (Figure 5c-bottom; slope -0.95% per option, 𝐹 = 34.66, 𝑝 = 0.00061) in the deterministic cases. Notably, to attain the performance of EDM, the theoretical maximizing agents required substantially more evaluations in situations involving a large number of options (Figure 5c gray; top: slope 2.8 evaluations per option, 𝐹 = 21.80, 𝑝 = 0.0023; bottom: slope 3.4 evaluations per option, 𝐹 = 250.2, 𝑝 = 9.7 × 10−7)."

3) The Discussion now includes a dedicated paragraph on the testability of the EDM predictions at the behavioral and neural levels:

"EDM is testable at the behavioral (Figure 5) and neural (Figure 6) levels. At the behavioral level, EDM possesses three distinctive characteristics. First, EDM obtains high reward rapidly (Figure 5a). This characteristic can be tested in choice environments that minimize noise as an additional factor (e.g., Figure 7a), providing performance versus evaluations plots analogous to Figure 2. Second, EDM allocates relative effort to relative utilities proportionally (Figure 5b). This characteristic can be tested in situations in which utilities can be measured precisely, e.g., through the volume of fluid rewards in animal experiments or money in human experiments. And third, EDM allocates effort rapidly regardless of the number of options. This is because the 𝑒harvest = 𝑢 (𝑒eval) strategy is agnostic to the number of options. This characteristic can be tested by varying the number of decision options and quantifying the number of times a decision-maker evaluates the options. At the neural level, EDM only requires the encoding of relative utilities of the recently sampled options. This relative code can be implemented using firing rates of the neuronal pools representing each alternative (Figure 6 top row, middle column). Indeed, this representation has been found in the primate brain. Specifically, the relative value associated with EDM, termed “fractional income,” captures firing rates of neurons in monkey area LIP (Sugrue et al., 2004)."

Reviewer #2 (Public Review):

In this article, Kubanek shows how simple, local decision strategies approximate optimal foraging behavior using analytical methods and model simulations. To ground the argument beyond model simulations, Kubanek generalizes previous theoretical frameworks in economics and foraging to show how evaluating relative utility and effort are sufficient to find optimal behavior. A particular strength of this study is its principled approach to linking general economic theory with foraging theory and deriving the conditions under which local behavioral strategies provide general and efficient means to the optimality problem. Re-casting utility and effort in relative terms offers attractive possibilities to apply these formulations in describing a range of phenomena. I, therefore, believe this short report will be of interest to a multidisciplinary audience from economics, psychology behavior theory, foraging, and neuroscience. The author's main claims are supported by their evidence.

Potential weaknesses of the study include:

1) Predictions from the proposed EDM framework are stated in vague terms and could be formulated more concretely and, if possible, included in the model simulations.

The predictions are now stated explicitly in a new Figure 5 and the associated text, and supported through performance evaluations within deterministic and stochastic choice environments.

Moreover, a new Figure 6 provides a representational and computational account of EDM, and summarizes the main point of the paper that EDM combines high performance with simple, biologically plausible evaluation.

2) The specificity of the EDM model and related model is only briefly touched on. The EDM argument could be strengthened by making the relation to other behavioral models more explicit.

The new Figure 5 now compares the key characteristics of EDM with a set of related and more complex models, and evaluates their performance across these characteristics. The relations to other behavioral models are further specified in two new Discussion paragraphs.

3) Many behavioral situations, including the in this paper often-cited study by Sugrue et al (2004), involve reward contingencies with a high level of uncertainty and non-stationary environments. While the author mentions these situations at the end of the discussion, it remains vague how EDM precisely performs or relates to decision strategies that deal with such environments.

The article now includes also stochastic choice environments, in addition to the original deterministic choice environments. This has resulted in new Figure 4, Figure 5, and the associated text.

The results in the stochastic environment corroborate those obtained in the deterministic environments.

Author Response

Reviewer #1 (Public Review):

Edmondson et al. develop an efficient coding approach to study resource allocation in resource constrained sensory systems, with a particular focus on somatosensory representations. Their approach is based on a simple, yet novel insight. Namely - to achieve output decorrelation when encoding stimuli from regions with different input statistics, neurons in the sensory bottleneck should be allocated to these regions according to jointly sorted eigenvalues of the input covariance matrix. The authors demonstrate that, even in a simple scenario, this allocation scheme leads to a complex, non-monotonic relationship between the number of neurons representing each region, receptor density and input statistics. To demonstrate the utility of their approach, the authors generate predictions about cortical representations in the star-nosed mole, and observe a close match between theory and data.

Strengths:

These results are certainly interesting and address an issue which to my knowledge has not been studied in-depth before. Touch is a sensory modality rarely mentioned in theoretical studies of sensory coding, and this work contributes to this direction of research.

A clear strength of the paper is that it demonstrates the existence of non-trivial dependence between resource allocation, bottleneck size and input statistics. Discussion of this relationship highlights the importance of nuance and subtlety in theoretical predictions in neuroscience.

The proposed theory can be applied to interpret experimental observations - as demonstrated with the example of the star-nosed mole. The prediction of cortical resource allocation is a close match to experimental data.

We thank the reviewer for the feedback. Indeed, demonstrating an ‘interesting’ effect in even such a simple model was one of the main aims.

Weaknesses:

The central weakness of this work are the strong assumptions which are not clearly stated. In result, the consequences of these assumptions are not discussed in sufficient depth which may limit the generality of the proposed approach. In particular:

1) The paper focuses on a setting with vanishing input noise, where the efficient coding strategy is toreduce the redundancy of the output (for example through decorrelation). This is fine, however, it is not a general efficient coding solution as indicated in the introduction - it is a specific scenario with concrete assumptions, which should be clearly discussed from the beginning.

2) The model assumes that the goal of the system is to generate outputs, whose covariance structure isan identity matrix (Eq. 1). This corresponds to three assumptions: a) variances of output neurons are equalized, b) the total amount of output variance is equal to M (i.e. the number of of output neurons), c) the activity of output neurons is decorrelated. The paper focuses only on the assumption c), and does not discuss consequences or biological plausibility of assumptions a) and b).

We have clarified the assumptions in the revised version. The original version did not distinguish clearly between assumptions that were necessary to allow study of the main effect, and assumptions that were included to present a full model but that could have been chosen otherwise without affecting the results.

This has now been made much clearer. Regarding the noise issue (point 1), we have clarified the main strategy pursued by the model namely decorrelation, we acknowledge other possible strategies, and we make clear whether and how noise could be incorporated into the model. Regarding the biological plausibility of our assumptions (point 2),

Reviewer #2 (Public Review):

The authors propose a new way of looking at the amount of cortical resources (neurons, synapses, and surface area) allocated to process information coming from multiple sensory areas. This is the first theoretical treatment of attempting to answer this question with the framework of efficient coding that states that information should be preserved as much as possible throughout the early sensory stages. This is especially important when there is an explicit bottleneck such that some information has to be discarded. In this current paper, the bottleneck is quantified as the number of dimensions in a continuous space. Using only the second-order statistics of the stimulus, and assuming only the second-order statistics carrying information, the authors use variance instead of Shannon's information. The result is a non-trivial analysis of ordering in the eigenvalues of the corresponding representations. Using clever mathematical approximations, the authors arrive at an analytical expression -- advantageous since numerical evaluation of this problem is tricky due to the long thin tails of the eigenvalues of the chosen covariance function (common in decaying translation-invariant covariances). By changing the relative stimulus power (activity ratio), receptor density (effectively the width of the covariance function), and the truncation of dimensions (bottleneck width), they show that the cortical allocation ratio, surprisingly, is a non-trivial function of such variables. There are a number of weaknesses in this approach, however, it produced valuable insights that have a potential to start a new field of studying such resource allocation problems all across different sensory systems in different animals.

Strengths

• A new application of the efficient coding framework to a neural resource allocation problem given acommon bottleneck for multiple independent input regions. It's an innovation (initial results presented at NeurIPS 2019) that brings normative theory with qualitative predictions that may shed new light to seemingly disproportionate cortical allocations. This problem did not have a normative treatment prior to this paper.

• New insights into allocation of encoding resources as a function of bottleneck, stimulus distribution, andreceptor density. The cortical allocation ratios have nontrivial relations that were not shown before.

• An analytical method for approximating ordered eigenvalues for a specific stimulus distribution.

Weaknesses

The analysis is limited to noiseless systems. This may be a good approximation in the high signal-to-noise ratio regime. However, since the analysis of allocation ratio is very sensitive to the tail of eigenvalue distribution (and their relative rank order), not all conclusions from the current analysis may be robust. Supplemental figure S5 perhaps paints a better picture since it defines the bottleneck as a function of total variance explained instead of number of dimensions. The non-monotonic nonlinear effects are indeed mostly in the last 10% or so of the total variance.

We agree that the model is most likely to apply in the low-noise regime, as stated in the Discussion. The robustness of the results is indeed a worry, and indeed we have encountered some difficulties when calculating model results numerically due to the issue pointed out by the reviewer, and this led us to focus on an analytical approach in the first case. However, to test model robustness we have now included numerical results for several other covariance functions to demonstrate that, at least qualitatively, the results presented in the paper are not simply a consequence of the particular correlation structure we investigated.

In case where the stimulus distribution is Guassian, the proposed covariance implies that the stimulus distribution is limited to spatial Gaussian processes with Ornstein-Uhlenbeck prior with two parameters: (inverse) length-scale and variance. While this special case allowed the authors to approach the problem analytically, it is not a widely used natural stimuli distribution as far as I know. This assumed covariance in the stimulus space is quite rough, i.e., each realization of the stimulus is spatially continuous isn't differentiable. In terms of texture, this corresponds to rough surfaces. Of course, if the stimulus distribution is not Gaussian, this may not be the case. However, the authors only described the distribution in terms of the covariance function, and lacks additional detail to fill in this gap.

We would argue that somewhat ‘rough’ covariance structure might be relatively common, for example in vision objects have clear borders leading to a power law relation and similarly in touch objects are either in contact with the skin or they are not. In either case, we have now extended the analysis to test several other covariance functions numerically. We found that, qualitatively, the main effects described in the paper were still present, though they could differ quantitatively. Interestingly, the convergence limit appeared to depend on the roughness/smoothness of the covariance function, indicating that this might be an important factor.

The neural response model is unrealistic: Neuronal responses are assumed to be continuous with arbitrary variance. Since the signal is carried by the variance in this manuscript, the resource allocation counts the linear dimensions that this arbitrary variance can be encoded in. Suppose there are 100 neurons that encode a single external variable, for example, a uniform pressure plate stimulus that matches the full range of each sensory receptor. For this stimulus statistics, the variance of all neurons can be combined to a single cortical neuron with 100 times the variance of a single receptor neuron. In this contrived example, the problem is that the cortical neuron can't physiologically have 100 times the variance of the sensory neuron. This study is lacking power constraint that most efficient coding frameworks have (e.g. Atick & Redlich 1990).

We agree that the response model, as presented, is very simplistic. However, the model can easily be extended to include a variety of constraints, including power constraints, without affecting the results at all. Unfortunately, we did not make this clear enough in the original version. The underlying reason is that decorrelation does not uniquely specify a linear transform and the remaining degrees of freedom can be used to enforce other constraints. As the allocation depends only on the decorrelation process (via PCA), we do not explicitly calculate receptive fields in the paper and any additional constraints (power, sparsity) would affect the receptive fields only and so were left out in the original specification. We have now added clearer pointers for how these could be included and why their inclusion would not affect the present results.

The star-nosed mole shows that the usage statistics (translated to activity ratio) better explains the cortical allocation than the receptor density. However, the evidence presented for the full model being better than either factor is weak.

We agree that the results do not present definitive evidence that the model directly accounts for cortical allocations and as we state in the paper, much stronger tests would be needed. Our idea here was to test whether, in principle, the model predictions are compatible with empirical evidence and therefore whether such models could become plausible candidates for explaining neural resource allocation problems. This seems to be the case, even though the evidence in favour of the ‘full model’ versus the ‘activity only’ model is indeed not overwhelming (though this might be expected as the regional differences in activity levels are much greater than those in density). We have now added additional tests to show that the results are not trivial. We would also like to note that it is not obvious that the ‘full’ model would perform better than the ‘activity only’ model: for either we choose the best-fitting bottleneck width (as the true bottleneck width is unknown), and therefore the degrees of freedom are equal (with both activity levels and densities fixed by empirical data).

Reviewer #3 (Public Review):

This work follows on a large body of work on efficient coding in sensory processing, but adds a novel angle: How do non-uniform receptor densities and non-uniform stimulus statistics affect the optimal sensory representation?

The authors start with the motivating example of fingers and tactile receptors, which is well chosen, as it is not overstudied in the efficient coding literature. However, the connection between their model and the example seems to break down after a few lines when the authors state that they treat individual regions as independent, and set the covariance terms to zero. For finger, e.g. that would seem highly implausible, because we typically grasp objects with more than one finger, so that they will be frequently coactivated.

Our aim was to take a first stab at a model that could theoretically account for neural resource allocation under changes in receptor density and activity levels, and by necessity this initial model is rather simple. Choosing a monotonically decreasing covariance function along with some other simplifications allowed us to quantify the most basic effects, and do so analytically. Any future work should take more complex scenarios into account. Regarding the sense of touch, we agree that the correlational structure of the receptor inputs will be more complex than assumed here, however, whether and how this would affect the results is less clear: Across all tactile experiences (not just grasps, but also single finger activities like typing), cross-finger correlations might not be large compared to intra-finger ones. Unfortunately, there is currently relatively little empirical data on this. That said, we agree with the broader point that complex correlational structure can be found in sensory systems and would need to be taken into account when efficiently representing this information.

The bottleneck model posited by the authors requires global connectivity as they implement the bottleneck simply by limiting the number of eigenvectors that are used. Thus, in their model, every receptor potentially needs to be connected with every bottleneck neuron. One could also imagine more localized connectivity schemes that would seem more physiologically plausible given the observed connectivity patterns between receptors and relay neurons (e.g. in LGN in the visual system). It would be very interesting to know how this affects the predictions of the theory.

We agree that the model in its current form is not biologically plausible. While individual receptive fields can be extremely localised, the initial allocation of neurons to regions we describe in the paper relies on a global PCA, and it is not clear how this might be arrived at in practice under biological constraints. However, our aim here was to specify a normative model that generates the optimal allocation and thereby answer what the brain should be doing under ideal circumstances. Future work should definitely ask whether and how these allocations might be worked out in practice and how biological constraints would affect the solutions.

The representation of the results in the figures is very dense and due to the complex interplay between various factors not easy to digest. This paper would benefit tremendously from an interactive component, where parameters of the model can be changed, and the resulting surfaces and curves are updated.

We have aimed to make the figures as clear as possible, but do appreciate that the results are relatively complex as they depend on multiple parameters. The code for re-creating the figures is available on Github (https://github.com/lauraredmondson/expansion_contraction_sensory_bottlenecks), making it easy to explore scenarios not described in the paper.

For parts of the manuscript, not all conclusions made by the authors seem to follow directly from the figures: For example, the authors interpret Fig. 3 as showing that activation ratio determines more strongly whether a sensory representation expands or contracts than density ratio. This is true for small bottlenecks, but for relatively generous ones it seems the other way around. The interpretation by the authors, however, fits better the next paragraph, where they argue that the sensory resources should be relatively constant across the lifespan of an animal, and only stimulus statistics adapt. However, there are notable exceptions - for example, in a drastic example zebrafish change their sensory layout of the retina completely between larvae and adult.

We have amended the text for this section in the paper to more closely reflect the conclusions that can be drawn from the figure. These are summarised below.

The purpose of Fig. 3B is to show that knowledge of the activation ratio provides more information about the possible regime of the bottleneck allocations. We cannot tell the magnitude of the expansion or contraction from this information alone, or where in the bottleneck the expansion or contraction would occur. Typically, when we know the activation ratio only, we can tell whether regions will be expanded or contracted or whether both occur over all bottleneck sizes. For a given activation ratio (for example, a = 1:2, as shown in the 3B), we know that the lower activation region can be either contracted only or both expanded and contracted over the course of the bottleneck. In this case, regardless of the density ratio, the lower activation region cannot be contracted only. Conversely, for any density ratio (see dashed horizontal line in Fig. 3B), allocations can be in any regime.

In the final part of the manuscript, the authors apply their framework to the star nosed mole model system, which has some interesting properties; in particular, relevant parameters seem to be known. Fitting to their interpretation of the modeling outcomes, they conclude that a model that only captures stimulus statistics suffices to model the observed cortical allocations. However, additional work is necessary to make this point convincingly.

We have now included a further supplementary figure panel providing more details on the fitting procedure and results for each model. Given that we fit over a wide range of bottleneck sizes, where allocations for each ray can vary widely (see Figure 6, supplement 1A), we tested an additional model to confirm that the model requires accurate empirical density and/or activation values for each ray to provide a good fit to cortical data. Here we randomise the values for the density and activation of each ray within the possible range of values for each. We find that with this randomisation of the values the model performs poorly on fitting even with a range of bottleneck sizes. This suggests that the model can only be fitted to the empirical cortical data when using the empirically measured values.

Yes, but without other platforms to go to (or the capacity to code own platform) the data, though owned, has little value. Thus the user remains dependent on platforms also in the web3 space

Now imagine, as a performance optimization, a lib where you can write in IMGUI style, so you write the same amount of code, but behind the scenes, the lib is managing things, so it's actually free to do all the retained mode heavy lifting.

We just invented declarative programming.

This describes one of the most pleasing hacks I've ever come across. I just now tracked it down and added it to my bookmarks. (Not sure why it wasn't already there.)

You could also conceive of going one step further. When your app (doesn't actually have to be a game, though admittedly it's much easier for you if it is) is compiled with tweak.h, it gives it the power to paint the source file on the screen—so you don't actually have to switch over to your text editor to save it, etc. Suppose you want to provide custom inputs like Bret Victor-style sliders for numeric values. You could edit it in your text editor, or you could derp around with it in-app. Tweaking the value in-app should of course both update it wrt the app runtime but also still write the file to disk, too, so if live reloading is turned on in your text editor, whatever changes you make inside the live process image gets synced out.

This needs to be cleaned up.

This is taking the wrong parts as fixed.

Reviewer #1 (Public Review):

This work addresses the problem of accurately estimating dynamics parameters in single particle tracking applications. The authors give a very extensive overview of the current problems and solutions while dealing with imaging of diffusive motion of subcellular particles and challenges that one faces while trying to estimate the main parameters of interest, such as diffusion constants. The authors properly address the issues with short trajectories, which are typical in practice and propose two advanced approached, which successfully deal with the mentioned shortcomings (short trajectories from which it is difficult to estimate parameters reliably, and the measurement errors that contaminate the input data). The proposed techniques are very interesting, and the way how those pure mathematical (and long existing) concepts are applied for this specific application of single particle tracking is rather novel. The proposed methodology is supported by a thorough validation, which includes simulations of all possible conditions (numbers of trajectories, distributions of the diffusion constants within the population of particles, the levels of inaccuracies in the measurements, etc.). Additionally, the experiments with the real data are also very convincing. The authors do focus on a regular Brownian diffusion and hopefully will show the applicability of these approaches to more typical applications containing anomalous diffusion. The availability of the code, which the authors provide on github, is very important, especially to less (technically) skilled audience from the field of experimental biology, who would like to apply those techniques to their data.

Reviewer #2 (Public Review):

This paper will be of interest to the cellular biologists who perform single-particle tracking experiments and develop new tracking methodologies. The authors investigate a new way of estimating an unknown number of diffusion states from short single-molecule trajectories. Ideas developed in the paper are likely to be used for further algorithm development. The authors give the users access to a repository on GitHub that contains comprehensive code that supports the paper.

Comments:

1. The authors claim in their abstract that their algorithm works on tracks with variable localization precision. However, in their simulation section, they investigate only a single localization precision value per diffusion state. This doesn't match the reality since intensity and background from molecules (and over the field) is never a single value, but follows a distribution. To make this simulation realistic the authors would need to generate datasets where the localization precision comes from a distribution with different parameters per diffusion coefficient (e.g. a different mean and variance).<br /> 2. In the introduction the authors claim that they also address the problem of having an unknown number of diffusion states. However, I don't think that the paper addresses this issue, because with their current approach it is not possible to classify trajectory segments into a state.

See Too DRY - The Grep Test.

But even allowing for grep is too lax, in my view. It's too high a cost. If I've got some file open and am looking at some Identifier X, I should be able to both deterministically and easily figure out exactly which file X lives in.

Violating this is what sucks about C's text-pasting anti-module system, and it's annoying that Go's "package" system ends up causing similar problems. I shouldn't have to have a whole-project view. I should be able to just follow my nose.

Ah! This remark highlights a fundamental difference in understanding between two camps, which I have been (painfully) aware of, but the source of this confusion has eluded me until only just right now. (Really, this is a source of frustration going back years.)

In one camp, the advice "don't use global variables" is a way of attacking a bunch of things endemic to their use, most notably unnecessary coupling to spooky state. In another camp "no global variables" is understood to mean literally that and taken no further—so you can have as much spookiness as you like, and so long as the value is not directly accessible (visible) from, say, another given piece of code appearing at the top-level ("global") context, as with the way i is bound to the activation record in this example but is not accessible outside the scope of getGetNext, then you're good.

That is, there are two aspects to variables: visibility and extent, and the first interpretation seeks to avoid the negative effects on both dimensions, while the second is satisfied by narrowly prohibiting direct visibility across boundaries.

I find the latter interpretation bizarre and completely at odds with the spirit of the exhortation for avoiding globals in the first place.

(What's worse is the the second interpretation usually goes hand in hand with the practice of making extensive use of closures, which because they are propped up as being closely associated with functions, then leads people to regretfully refer to this style as functional programming. This is a grave error—and, to repeat, totally at odds with the spirit of the thing.)

All of the apparatus that Wiener designs have flaws that are based around human behavior. I think it would benefit him to learn about people's behaviors in a way that could explain them to computer code. I also believe that he sees cybernetics as a shortcut in this venture. Humans understand and can figure out human behaviors, we are good at identifying patterns, in a way that programs and non-humans are not. Ideally, the non-human part of a cybernetic being would bring tactics, consistency, and the ability to process information in an objective way. As where the human side would bring the ability to identify behaviors, patterns, and use senses to be processed by the non-human.

Author Response