Completed on 28 Jul 2016 by Patrick Mineault . Sourced from http://biorxiv.org/content/early/2016/12/07/045336.
Login to endorse this review.
Author response is in blue.
I appreciate the very narrow point that the authors are making - i.e. that a certain class of GLM-like models doesn't work very well when it comes to explaining RGC responses to natural stimuli -, but I feel like the claim that "retinal signaling
under natural conditions cannot be captured by models that begin with linear filtering" is a bit too broad.
First, a very simple two subunit-model qualifies as a "model which begins with linear filter"; e.g. two linear subunits, followed by a pointwise nonlinearity, then combined linearly and followed by another pointwise nonlinearity. Dan Butts (McFarland at al. 2011, J Neurosci) showed that that goes a long way in explaining variability in LGN cells in response to natural stimuli (http://neurotheory.umd.edu/Pub..., and one would presume would resolve some of the deficiencies in the RGC models.
Second, the claim that '[these findings] emphasize the importance of additional spatial nonlinearities, gain control, and/or peripheral effects in the first stage of visual processing." is both a bit strong and a lost opportunity. It makes it sound like GLMs and GLM+s are so deficient that you're going to need to throw the kitchen sink at the problem to explain natural image responses. Why not fit a simple subunit model and see if that helps? Then that will help us understand the limits of the current model and how much extra modeling really needs to be done to explain natural movie responses. The code is available for this (http://neurotheory.umd.edu/nim..., it seems like a pretty low-hanging fruit.
Thanks for these excellent comments and suggestions. I agree with the first comment entirely. The writing and interpretation will be improved in the next version. As to the second comment, a subunit model (if we are talking about the same thing) is a form of spatial nonlinearity. Indeed, it's the form that I was thinking of when I wrote that, and we have been developing approaches to fitting subunit models. We suspect that gain control will also be necessary, but I also lean toward the more parsimonious models whenever possible. We appreciate and will definitely take a look at your code.