On 2016 Oct 13, Lydia Maniatis commented:

In addition to the circularity of the noise revelation, the claim in the abstract regarding "spatial pooling" is also misleading:

" Our results suggest that chickens use spatial pooling of cone outputs to mitigate photon-shot noise."

The results show nothing of the sort, since "spatial pooling" was introduced as a free parameter to aid in data-fitting. If we were to label this parameter "the tooth fairy" then the data would "suggest" that the tooth fairy mediates color discrimination.

Again, we are talking about cosmetic applications of concepts to data-fitting techniques and arguments consisting of multiple casual assumptions that license no such interpretation.

On 2016 Oct 05, Lydia Maniatis commented:

Thanks for your reply, Peter. I've copied it below and respond to the various points in the order they came up.

You: There are measurements of noise in the visual system of animals which we have cited in the paper, so I can not agree with you that there is no evidence for noise in the visual systems of animals.

Me: I think that all of the cited papers assume they are measuring "noise", without actually testing that assumption.

You: In this paper we are describing the smallest colour difference that chickens were able to discriminate to figure out the equivalent Weber fraction which describe these limits. Whether that is actually caused by noise or not we can not say from our experiment, but it is adressed in some of the cited literature.

Me: The title of your paper says "behavioural thresholds reveal receptor noise" and the abstract states that your "experiments allowed us to compare behavioural results wtih model expectations and determine how different noise types limit colour discrimination." So it sounds like a claim that your experiments did, in fact, corroborate the "noise" assumptions. In fact, the noise argument was never in any danger of falsification, because it was not tested, only assumed to be indirectly measurable. The confusion about this issue has seeped into other publications; the reason I was looking at yours was that Scholtyssek, Osorio and Baddeley (2016) claim that your paper "validated experimentally" the "receptor-noise limited model of Vorobyev and Osorio (1998)." This didn't seem likely, so I wanted to check.

You: We find, similar to other experiments, that there this Weber fraction is the same in relatively brighter light (thresholds are very similar) but that the Weber fraction we need to assume to describe the thresholds needs to be higher in dim light, consistent with addition of photon-shot noise.

Me: "consistent with" doesn't mean causally connected to. If the claim is about causality, this needs to be tested, not assumed. The adjustments and assumptions in general seem ad hoc.

You: Which has been shown to limit visual sensitivity in many experiments.

I think this applies in very, very low light conditions. But your experiment does not constitute a test that this is relevant to your results. Again, it is simply assumed, even though the idea that raw responses of neurons early in the hierarchy are transmitted directly to visual experience is not credible.

You: We are using the Receptor Noise Limited model to describe the data, it is a tool that can be used to make predictions in other scenarios as well.

Me: Why do you consider it a useful tool? Have its "noise" related assumptions ever been tested and corroborated? How? When Vorobyev and Osorio (1998) say that the model "predicted" some psychophysical data, they just mean that they were able to find some datasets that fit more or less, and some that didn't, not that they predicted beforehand which datasets would match. And again, general consistency alone does not justify uncritical acceptance of whatever qualitative assumptions are tacked onto the math.

You: Regarding Percepts: We are not talking about how the animals perceive the colours, but rather how they are able to discriminate them. Absolutely, there is much more in the visual system going on that we yet do not understand. But the photoreceptor cells are the input of the system and the limits of the photoreceptors are going to be important.

Me: First, discrimination implies perception, at least to the extent that two surfaces are perceived as same or different. Second, the limits of photoreceptors are not discernible in ordinary percepts. Correct me if I'm wrong, but you don't even seem to be acknowledging the complexity of the neural code for color, which involves combining cone activity to produce color experiences that can't be qualitatively analyzed on the basis of the individual cone activity, for example the fact that "red" plus "green" plus "blue" cone activity (or, e.g. "red plus green") produces the experience of grey or white. With four cones, the code will be even more complicated.

On 2016 Oct 05, Peter Olsson commented:

Dear Lydia Thank you for your interest and comments.

There are measurements of noise in the visual system of animals which we have cited in the paper, so I can not agree with you that there is no evidence for noise in the visual systems of animals.

In this paper we are describing the smallest colour difference that chickens were able to discriminate to figure out the equivalent Weber fraction which describe these limits. Whether that is actually caused by noise or not we can not say from our experiment, but it is adressed in some of the cited literature.

We find, similar to other experiments, that there this Weber fraction is the same in relatively brighter light (thresholds are very similar) but that the Weber fraction we need to assume to describe the thresholds needs to be higher in dim light, consistent with addition of photon-shot noise. Which has been shown to limit visual sensitivity in many experiments. We are using the Receptor Noise Limited model to describe the data, it is a tool that can be used to make predictions in other scenarios as well.

Regarding Percepts: We are not talking about how the animals perceive the colours, but rather how they are able to discriminate them. Absolutely, there is much more in the visual system going on that we yet do not understand. But the photoreceptor cells are the input of the system and the limits of the photoreceptors are going to be important.

On 2016 Oct 01, Lydia Maniatis commented:

From the article (I've broken up a single continuous paragraph for clarity (caps mine):

"We ASSUME that discrimination thresholds of visual systems are set by noise (Vorobyev and Osorio, 1998; Lind and Kelber, 2009b) and that noise can arise from different sources in different light conditions. Over a wide range of relatively high light intensities...

...we EXPECT that Weber's law holds, so that sensitivity changes proportionally to light intensity and a constant Weber fraction (थscribes the signal-to-noise ratio that sets discrimination thresholds (Donner et al., 1990; Osorio et al, 2004; Lind et al., 2014). Under these conditions...

...the main source of noise is PROBABLY transducer noise (Lillywhite and Laughlin, 1979; Howard and Snyder, 1983) originating at the later stages of signal transduction in the photoreceptors, possibly by fluctuations in cGMP levels (Angueyra and Rieke, 2013).

There are no electrophysiological measurements of noise in bird photoreceptors, but rough estimates of noise have been deduced from the results of behavioural experiments on spectral sensitivity (Maier, 1992; Vorobyev et al., 1998; Lind et al., 2014)."

There is no empirical evidence that there is "noise" in the visual system and that this noise affects perception. The assumption is, furthermore, so vague that it could not be tested; and in our experience as observers, percepts certainly aren't "noisy." The reference to deduction of "rough estimates of noise" is contingent on this same untested, untestable, and empirically implausible assumption.

In addition, the idea that the formed, conscious percept, the product of complex processing with outcomes that resemble inferential logic, (such as, for example, lightness constancy (which has been shown even in chicks)) can be used to detect the neural properties of particular sets of neurons (here photoreceptors) is untenable.

Explicitly, the authors assume that "the discrimination thresholds are set by photoreceptor noise, which is propagated into higher order processing." This is a huge assumption, given what is known about the very different receptive field properties of neurons as we go up the hierarchy, and in view of the fact that these higher levels affect the activity of lower levels via feedback. It implies, in effect, that we can ignore, suppress or control for the participation of all other neural populations and interactions in this system. It is a claim that clearly requires a bit more effort at argument, rather than the casual assumptions, expectations, references to "established" or "standard" methods, etc. being offered here.

I see the bun, the catsup, the mustard and the onion rings...but where's the beef?

Another fascinating though not original aspect of this article is that it assumes that if we fit two arbitrary "models" to a dataset, and one of them fits this dataset better than the other, then the former is "more probable" than the other. This is the "Bayesian" way (though I don't think Bayes himself would agree). The fact that both models are too vague even to test doesn't interfere with these probabilities, which are in any case "subjective" and thus reflect the authors' (and the reviewers and editor's, I guess) personal beliefs.

On 2016 Oct 01, Lydia Maniatis commented:

From the article (I've broken up a single continuous paragraph for clarity (caps mine):

"We ASSUME that discrimination thresholds of visual systems are set by noise (Vorobyev and Osorio, 1998; Lind and Kelber, 2009b) and that noise can arise from different sources in different light conditions. Over a wide range of relatively high light intensities...

...we EXPECT that Weber's law holds, so that sensitivity changes proportionally to light intensity and a constant Weber fraction (थscribes the signal-to-noise ratio that sets discrimination thresholds (Donner et al., 1990; Osorio et al, 2004; Lind et al., 2014). Under these conditions...

...the main source of noise is PROBABLY transducer noise (Lillywhite and Laughlin, 1979; Howard and Snyder, 1983) originating at the later stages of signal transduction in the photoreceptors, possibly by fluctuations in cGMP levels (Angueyra and Rieke, 2013).

There are no electrophysiological measurements of noise in bird photoreceptors, but rough estimates of noise have been deduced from the results of behavioural experiments on spectral sensitivity (Maier, 1992; Vorobyev et al., 1998; Lind et al., 2014)."

There is no empirical evidence that there is "noise" in the visual system and that this noise affects perception. The assumption is, furthermore, so vague that it could not be tested; and in our experience as observers, percepts certainly aren't "noisy." The reference to deduction of "rough estimates of noise" is contingent on this same untested, untestable, and empirically implausible assumption.

In addition, the idea that the formed, conscious percept, the product of complex processing with outcomes that resemble inferential logic, (such as, for example, lightness constancy (which has been shown even in chicks)) can be used to detect the neural properties of particular sets of neurons (here photoreceptors) is untenable.

Explicitly, the authors assume that "the discrimination thresholds are set by photoreceptor noise, which is propagated into higher order processing." This is a huge assumption, given what is known about the very different receptive field properties of neurons as we go up the hierarchy, and in view of the fact that these higher levels affect the activity of lower levels via feedback. It implies, in effect, that we can ignore, suppress or control for the participation of all other neural populations and interactions in this system. It is a claim that clearly requires a bit more effort at argument, rather than the casual assumptions, expectations, references to "established" or "standard" methods, etc. being offered here.

I see the bun, the catsup, the mustard and the onion rings...but where's the beef?

Another fascinating though not original aspect of this article is that it assumes that if we fit two arbitrary "models" to a dataset, and one of them fits this dataset better than the other, then the former is "more probable" than the other. This is the "Bayesian" way (though I don't think Bayes himself would agree). The fact that both models are too vague even to test doesn't interfere with these probabilities, which are in any case "subjective" and thus reflect the authors' (and the reviewers and editor's, I guess) personal beliefs.

On 2016 Oct 13, Lydia Maniatis commented:

In addition to the circularity of the noise revelation, the claim in the abstract regarding "spatial pooling" is also misleading:

" Our results suggest that chickens use spatial pooling of cone outputs to mitigate photon-shot noise."

The results show nothing of the sort, since "spatial pooling" was introduced as a free parameter to aid in data-fitting. If we were to label this parameter "the tooth fairy" then the data would "suggest" that the tooth fairy mediates color discrimination.

Again, we are talking about cosmetic applications of concepts to data-fitting techniques and arguments consisting of multiple casual assumptions that license no such interpretation.