we found that prior statistics warp neural representations in the frontal cortex allowing the mapping of sensory inputs to motor outputs to incorporate prior statistics in accordance with Bayesian inference. Analysis of recurrent neural network models performing the task revealed that this warping was enabled by a low-dimensional curved manifold, and allowed us to further probe the potential causal underpinnings of this computational strategy. These results uncover a simple and general principle whereby prior beliefs exert their influence on behavior by sculpting cortical latent dynamics.

we observed transient increases in neural activity in both direct- and indirect-pathway SPNs when animals initiated actions, but not when they were inactive

This normative model provides a core insight: the “optimal subspace” is likely high dimensional, with many different initial conditions giving rise to the same correct movement.

‘elemental affect’ [51] – a representationalsignature that signals to downstream information processing systems that the incoming perceptualstimulus is somehow distinct from other stimuli, and therefore worthy of further attention.

Modern perceptual machines are in fact notoriously limited in terms of abstraction

features most predictive of affect are the deepest features

object and scene recognition

arousal

Such a stable representation is a hallmarkof the reverberating activity constituting NCCs in the GNW the-ory (Dehaene et al. 2003) and, interestingly, was reduced in theRSVP condition

we predict thatexperiments 3[3], 4[3], and 4[4] performed in the general setting –i.e. with integral dimensions, with children or monkeys as subjects,or when categorization difficulty is extrapolated from errors inidentification learning – should be better explained by ˆumean thanˆumin

n Feldman (2000,2003), he identified the extended catalog of logical categorystructures beyond SH

That is, some subjects learn Type II veryfast and some learn it quite slow, and aggregate results essentiallyreflect the weighting of the two types of learners in a sample. Itmay be the case that subjects who are, implicitly or explicitly, con-structing rules that focus on fixing one dimension are the ones wholearn Type II slowly, but those who are constructing rules that fo-cus on fixing two dimensions are the ones who learn it quickly.It also might be that agents who focus on one rule over the otherweight that information more heavily:

This is plausibly well-defined, in that the Boolean complexity startswith a maximally redundant list of the elements of a categoryand then reduces redundant elements: perhaps the complexityvalues at each step could be averaged

Part of the distinction is thatwhile GIST searches for ‘invariants’, it makes no distinction amongthe non-invariants about their degree of variance, while our metricdoes.

Now consider d = 2. If two dimensions are known,what is the uncertainty about the stimulus category A? First, con-sider the first two dimensions. When the first two dimensions areknown, then the category is known with certainty. That impliesthat the uncertainty associated with the first two dimensions iszero. Now consider the case when the first and third dimensionsare known. In this case, the category is completely unknown (thereas many stimuli with first and third dimensions equaling, for ex-ample, (0, 0), in the category A as B) and so uncertainty is maxi-mized at 1. This same argument applies when the second and thirddimensions are known (uncertainty is maximized).

that is, there are as many cat-egory A stimuli with a zero as the first entry as category B stimuli,and so on

metric

Table 2

this applies to the paradigm-specific order, in which so-phisticated learners can observe separable dimensions and mayemploy abstraction or attention with regard to these dimensions

The core finding is that Type II was moredifficult for the monkeys to learn than Types III–V (which the au-thors elect to average across in their reporting)

SHJ types

There arefour separate cases that yield results consistent with this order-ing: first, stimulus generalization theory, which generates a pre-diction of the ordering of the classification problems based on thefrequency of mistakes (pairwise confusions) in learning uniquelabels (i.e., identification learning) for each item (Shepard et al.,1961); second, stimuli comprised of integral dimensions (Garner,1974) that are difficult for the learner to perceptually analyze anddistinguish, such as brightness, hue, and saturation (Nosofsky &Palmeri, 1996); third, learning by monkeys (Smith, Minda, & Wash-burn, 2004); fourth, learning by children (Minda, Desroches, &Church, 2008).

Our results imply that redundancy across a distributed network could mask possible causalroles in optogenetics experiments

This is inconsistent withattractors with a pair of fixed points (one for each choice condition) 29

Remarkably, behavioral performance was unaffected (Fig. 5b, control trials, before vs. aftercallosotomy, p>0.05, two-tailed t-test), with normal performance 17 hours after callosotomy(Extended Data Fig. 10b). However, behavioral performance was now highly sensitive totransient unilateral photoinhibition

Circuit dynamics are actively restored alongbehavior-related directions in the activity space, but not along certain non-informativedirections

his ‘ramping mode’ could reflect non-specific‘urgency’ 39 driven by a source external to ALM.

Second, we obtained a mode that maximized sustained effects of ipsilateral perturbations(‘persistent mode’, Fig. 3d). By construction, the persistent mode was altered by theperturbation, up to and beyond movement onset. However, this projection did notdiscriminate trial type nor predict behavior on control trials

that activity “caught up” to reach the same level as in unperturbed trials

Transient (duration, 0.5 s) unilateralphotoinhibition of ALM up to the ‘go’ cue (late delay) caused an ipsilateral response bias (n= 5 mice, p < 0.01, two-tailed t-test; Fig. 1d), similar in magnitude to photoinhibition overthe entire delay epoch (Extended Data Fig. 3) 3, 30. In contrast, photoinhibition ending atleast 0.3 s prior to the ‘go’ cue produced minimal behavioral effects (middle delay, earlydelay; p > 0.1, two-tailed t-test).

Models of persistent and ramping activity 5, 18, 22-24, 35-37 do not recover after transientlysilencing subsets of neurons (Fig. 1c; Extended Data Fig. 1f-i). We transiently silencedpreparatory activity

Network models can produce persistent and ramping activity, including integrators 15, 18-21and trained recurrent networks 22-24.

Perturbations to one hemisphere are thus corrected by informationfrom the other hemisphere.

detailed neural dynamics that drive specific future movementswere quickly and selectively restored by the network

In this condition, feedback was delayed by 2.5 seconds after the response

Second, note that the two-stage model learns consistently, both during training and during transfer. Thus, this model successfully solves the credit assignment problem.

distinguished between an associative loop through the caudate nucleus and a motor loop through the putamen

One clue comes from fMRI studies of II category learning, which have not reported a consistent site of task-related activation within the striatum. Some studies have reported activation in both the caudate nucleus and the putamen (Cincotta & Seger 2007; Seger & Cincotta, 2002), while other studies have reported task-related activity either only in the putamen

In contrast, creating new categories from the same stimuli in any other way should be less disruptive, because only some of the associations would have to be relearned, but not all. Existing empirical data, however, indicate that reversal learning is easier than learning novel categories

recovery from a full reversal should be easier than learning new categories equated for difficulty

D go

D

thal ALM

Inputs to ALM-projecting thalamus

For example, preparatory activ-ity emerged in PTlower neurons during the delay epoch (along CDlate)and persisted through the go cue and up to the termination of lickingbouts. In the same cell type, and sometimes in the same individualcells (for example, Fig. 6d; cell #3), activity was modulated after thego cue along a different direction (CDgo), consistent with a movementcommand.

The thalamus also receivesa projection from L6 corticothalamic neurons, but these neurons aresparsely active, uncoupled from PT neurons, and have weak synapseson thalamic neurons

Previous studies have suggested that collaterals of PT neurons to thethalamus13,14 might provide an efference copy of motor commands

terminating movements

PTlower population revealed neurons that were strongly modu-lated at the go cue, at the offset of movement, or both

In this projection, selectivitywas larger and more consistent across trials in the PTupper populationcompared to the PTlower population (Fig. 5b). Furthermore, selectivityin the PTupper population remained constant throughout the sample anddelay epochs until the go cue

Trial-averaged activity patterns were nearly as diverse within each PTpopulation as across all neurons recorded within ALM

Unlike previous observations in layer 2/3, activity accompanying learned movements did not become more consistent with learning; instead, the activity of dissimilar movements became more decorrelated. These results indicate that the relationship between corticospinal activity and movement is dynamic and that the types of activity and plasticity are different from and possibly complementary to those in layer 2/3. <div class="c-nature-box c-nature-box--side " data-component="entitlement-box"> <p class="c-nature-box__text js-text">You have full access to this article via your institution.</p> <div class="c-pdf-download u-clear-both js-pdf-download"> <a href="/articles/nn.4596.pdf" class="u-button u-button--full-width u-button--primary u-justify-content-space-between c-pdf-download__link" data-article-pdf="true" data-readcube-pdf-url="true" data-test="download-pdf" data-draft-ignore="true" data-track="click" data-track-action="download pdf" data-track-label="link" data-track-external download> <span class="c-pdf-download__text">Download PDF</span> <svg aria-hidden="true" focusable="false" width="16" height="16" class="u-icon"><use xlink:href="#icon-download"/></svg> </a> </div> </div> You have full access to this article via CUNY Office of Graduate Studies Download PDF

A classical way to probecortical output potential has been to monitor behavioral pat-terns elicited by microstimulation of cortical regions

notably its most anterior (referred to as anterior lateralmotor cortex [ALM]) and lateral domain (referred to as tongue/jaw motor cortex [tjM1]) were previously studied in the contextof orofacial behaviors including licking

other subcortical structures

Distinct cortical regions generate three-dimensional syn-aptic columns tiling the lateral medulla, topographically matching the dorso-ventral positions of postsynapticneurons tuned to distinct forelimb action phases.

uperior colliculus total: 97,Slco2a1: 43, Npsr1: 15, Hpgd: 39; pons total: 100, Slco2a1: 86, Npsr1: 5,Hpgd: 9; Fig. 2

Similarly, thalamus-projecting PT neurons mapped to the Npsr1 andHpgd clusters

PT neurons form the sole cortical pro-jection to motor areas in the midbrain and hindbrain

These results indicate that two types of motor cortexoutput neurons have specialized roles in motor control

Intratelencephalic neurons in layers (L) 2–6 receive input fromother cortical areas and excite pyramidal tract (PT) neurons5–7 . PTneurons, the somata of which define neocortical L5b 8, link the motorcortex with premotor centres in the brainstem and spinal cord9,10 anddirectly influence behaviour10–12. PT neurons also project to the thala-mus6,13,14. Preparatory activity requires reverberations in a thalamocor-tical loop15. Consistent with roles in both the planning and initiation ofmovement, PT neurons are structurally heterogeneous14,16,17 and showdiverse activity patterns, including preparatory activity and movementcommands18–20.

motor cortex population activity to aninitial condition

they constitute a learned response thatis specific to the sound used as the Go cue

Although PT lowercells have only weak connections with other pyramidal cells(Brown and Hestrin, 2009; Kiritani et al., 2012), they may influ-ence the network via their connections to local GABAergic inter-neurons or through multi-regional loops

GtACR1

Our hypothesis predicts that although both D go and CDresponseappear after the Go cue, they may be dissociable with manipula-tion of the CD response

Activity along Dgo is non-selective(Figure 1E) and cannot decode lick direction (Figure S1J) becauseactivity changes around the Go cue are largely similar across trial

Activity along Dgo explains a largeproportion of ALM activity after the Go cue (Figure S1I), similar tothe ‘‘condition-invariant signal’’ described in a primate reachingtask

This finding is consistent with theobservation that fine-scale movement parameters and reactiontimes are coded in preparatory activity (unpublished observa-tions) (Li et al., 2016; Even-Chen et al., 2019) and implies thatALM preparatory activity (activity along CD delay) contributes tocontrol of future movements (activity along CD response

G

These two modes together explain71.2 (65.3–76.0) % (mean, 2.5%–97.5%confidence interval) of selectivity in ALMaround the movement initiation

Pearson’s correlation of this population selec-tivity vector is high across time within the delay epoch (Figure 1D,a box with white dotted outline), implying that a similar combina-tion of ALM neurons maintains selectivity during motor planning

e defined a populationselectivity vector: wt = rlick-right, t
rlick-left, t, where rlick-right, tand rlick-left, t are vectors of spike rate of individual neurons foreach time t, averaged over lick right and left trials, respectively(the number of elements in the vector equals the number of re-corded neurons)

In addition, a subset of cells (177/5,136 cells)switched selectivity.

These modes occupy near-orthogonal subspaces

Another activitymode after the Go cue consists of changes that are invariant to themovement type (condition-invariant signal;

collapses

These trajectories are typically confined to a low-dimensionalmanifold, defined by several ‘‘activity modes’’ that explain a sig-nificant proportion of the population activity. Activity modes canbe obtained by projecting neural activity along specific direc-tions in neural state space, or similar dimensionality reductionmethods

This ‘‘pre-paratory activity’’ encodes specific upcoming movements,often seconds before movement onset

DA-Mediated Signaling in Activation of Mitogen-Activated Protein Kinases

sequential activity

Thethalamus-projecting PT neurons maintain planning-related activity during the memory epoch. Incontrast, the medulla-projecting neurons develop selectivity late and also carry a large proportionof the execution mode after the Go cue

Two-photon-mediatedoptogenetic stimulation of a small number (<10) of ALM neurons revealed sparse subnetworksthat independently maintain activity and are only weakly coupled to other subnetworks (Daie

Remarkably, bilateral silencing of large regions of cortex posterior to ALM, including pri-mary motor cortex, sensory cortex, and parietal cortex, has little effect on preparatory activity andsubsequent behavio

he same input from the sensory cortex that drives decision-makingand preparatory activity in ALM loses influence

ALM

near-orthogonal

In a memory-guided movement task in mice, distractors early in thememory epoch, or on trials with low ramping, biased choices by shifting ALM dynamics, whereasdistractors late in the delay, or on trials with high ramping, did not affect choice or ALM dynam-ics (Finkelstein et al. 2021)

Second, training a recurrent neural network to mimic ALM activity patterns and their responseto perturbations requires an external ramping input

Figure 3

For example, can a single network multiplexmultiple attractor landscapes to perform multiple tasks (Gallego et al. 2018, Yang et al. 2019)? Howis the attractor landscape shaped by learning (Sadtler et al. 2014, Sun et al. 2022)? And how doesthe sensory information feed into the attractor landscape?

Continuous attractor model

Following a perturbation, a continuousattractor will maintain a trace of the perturbation, corresponding to a displacement in activity

ehavioral effects of the perturbations have been used to test algorithmic models. For example,silencing FOF biases choice in a perceptual decision-making task (Erlich et al. 2011, Piet et al.2017), consistent with models in which FOF maintains the binary choice (motor plan) with pointattractors but not with FOF making decisions with continuous attractors

Modes are par-ticular directions in these subspaces and are typically chosen to reveal interpretable features of thedata. For example, in a licking task, projections along a vector in activity space that maximally dis-tinguish movement directions contain nearly all licking direction-selective activity (Figure 1c,d)

whereas continuous attrac-tors allow integration and storage of continuous variables

attractor memory system resembles a ball rolling in a hilly landscape

The network mechanisms underlying preparatory activity and the transitionto another state (e.g., movement initiation) are of great interest because similar mechanisms mightunderlie diverse cognitive functions, including working memory and integration of evidence forsensory decision-making

t is unknown whether preparatory activity causes idiosyncraticmicromovements, movement contributes to preparatory activity, or preparatory activity and movement are modu-lated by a common input.

Neural dynamics in cortex causes muscle tension, and sensory feedback modulatescortical activity in turn

Preparatory activ-ity is thought to set the state of neural activity to initial conditions that favor accurate and rapidmovements, with different initial conditions corresponding to different movements

have time to plan

The serotonin system also plays an important role in anesthesia response. Inhalational anesthetic agents may work in part by suppressing serotonin release, and patients taking serotonergic antidepressants may require increased dosage of these agents (65).

Thus, DAopto-induced PKA activation must besuppressed in distal dendrites in order to attain the large dynamic range for the timingdetection

proximal (first branch) dendrites

AC1 blocker (NB001)

Unlike structural plasticity or Camuiα-CR activation, PKA activation in response tostimulation of a dendritic spine by STDP and DAopto was not restricted to the stimulatedspine; neighboring spines also exhibited significant PKA activation

). We found that DAoptodid not affect Ca2+ transients (Fig. 3, B to D, and fig. S6, A to C), indicating that Ca2+signaling modulation did not play a major role.

high phosphodiesteraseactivity

the back-propagating action potential was attenuated, so that the dopamine timing effect was small.

In the both down- and up-states, the effects of dopamine inputs were the most prominent when DA preceded Post by approximately 60 ms, and amplified more when Glu preceded Post.

Model prediction of voltage and calcium responses to bAP in dendritic spines.

The voltage and calcium responses to a back-propagating postsynaptic action potential (bAP)

dopamine enhances KIR, Cav1.2 and NMDAR conductances and reduces NaF, CaN and CaQ conductances in the striatum,