Sensory coding and contrast invariance emerge from the control of plastic inhibition over emergent selectivity
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
Visual stimuli are represented by a highly efficient code in the primary visual cortex, but the development of this code is still unclear. Two distinct factors control coding efficiency: Representational efficiency, which is determined by neuronal tuning diversity, and metabolic efficiency, which is influenced by neuronal gain. How these determinants of coding efficiency are shaped during development, supported by excitatory and inhibitory plasticity, is only partially understood. We investigate a fully plastic spiking network of the primary visual cortex, building on phenomenological plasticity rules. Our results suggest that inhibitory plasticity is key to the emergence of tuning diversity and accurate input encoding. We show that inhibitory feedback (random and specific) increases the metabolic efficiency by implementing a gain control mechanism. Interestingly, this led to the spontaneous emergence of contrast-invariant tuning curves. Our findings highlight that (1) interneuron plasticity is key to the development of tuning diversity and (2) that efficient sensory representations are an emergent property of the resulting network.
Article activity feed
-
###Reviewer #3:
The author implemented a recurrent network with excitatory plasticity (from Clopath10) and inhibitory plasticity (from Vogels11) at all connections - both feedforward and recurrent. They showed that a model with inhibitory plasticity exhibits more diverse receptive fields (covering the different orientation preferences more uniformly) compared to a model without any inhibition but with plastic excitatory synapses. They showed that synaptic connectivity reflects tuning similarity. They then showed that inhibition helps decorrelation. In their model, inhibition sharpens tuning curves and helps to exhibit contrast invariance as well as promotes sparseness. Finally, they showed that their plastic model has a lower reconstruction error compared to a model without inhibition at all but similar to a model where inhibition is …
###Reviewer #3:
The author implemented a recurrent network with excitatory plasticity (from Clopath10) and inhibitory plasticity (from Vogels11) at all connections - both feedforward and recurrent. They showed that a model with inhibitory plasticity exhibits more diverse receptive fields (covering the different orientation preferences more uniformly) compared to a model without any inhibition but with plastic excitatory synapses. They showed that synaptic connectivity reflects tuning similarity. They then showed that inhibition helps decorrelation. In their model, inhibition sharpens tuning curves and helps to exhibit contrast invariance as well as promotes sparseness. Finally, they showed that their plastic model has a lower reconstruction error compared to a model without inhibition at all but similar to a model where inhibition is blocked after learning.
Below is a list of questions/comments:
The finding regarding receptive field diversity is probably the most novel part of the paper. It would be nice to dig into it a bit more. Does inhibitory plasticity or inhibition promote receptive field diversity? And what is the intuition behind it? Why?
It would be good to discuss the various histograms of orientation preference reported in different experimental data and compare that to the model.
The introductory paragraph of the results section does not contain enough information to understand the results. Without reading the Methods first, it is very confusing. In particular:
-The 2:1 and 3:1 model variants are poorly explained. This comes from the different levels of \rho but how it is written, it seems to come from a difference in connectivity or the ratio between the numbers of E and I cells.
-Noihn model: it should be noted that excitation is plastic.
The authors report the correlation drop with and without inhibition (l120-130). Would it be possible to compare quantitatively to some experimental data where inhibition is blocked (e.g. optogenetically). And so, how much does this drop depend on the model parameters?
Plasticity inhibition helps reconstruction error. It would be nice to elaborate further. In Fig 9a, surprisingly blockInh is doing very well. Why? I am not sure the statements in the text (regarding the role of inhibitory plasticity on the reconstruction error and encoding quality) are supported by the simulation results.
I encourage the author to be more precise in the text: what comes from inhibition, which effect can you get with fixed inhibition (tuned or broad), what comes from plasticity inhibition, what has been shown before etc. For example, I compile a little list below that helps me putting things together:
-Fig 3. synaptic connectivity reflects tuning similarity - Shown in Clopath10
-Fig 4: Inhibitory strength influence the response decorrelation- Shown in Vogels11
-Fig 5: Inhibition sharpens tuning curves - that's the classical iceberg effect. It works with fixed blanket of inhibition - e.g. Ben-Yishai 95.
-Fig 6-7. Inhibition leads to contrast invariance. Same here, inhibition does not need to be plasticity, it works with blanket inhibition - e.g. Ben-Yishai 95.
-Fig 8. Inhibition increases sparseness - Vogels11 inhibition plasticity leads to E/I balance with increased sparseness.
- The code should be made public.
-
###Reviewer #2:
The authors introduce a computational model of the interplay between excitatory and inhibitory plasticity during development in V1. The analysis of the work is interesting; however, several assumptions have to be checked and a multitude of additional analyses is required to validate the conclusions.
Major Comments:
The model describes the dynamics during the development of V1. However, during development there are several phases, each having its specific properties and dynamics. For instance, van Versendaal and Levelt 2016 discuss that especially inhibition could have a critical and phase-specific role. Please discuss in more detail the relation of the model to the developmental periods or rather which period you model.
In the model, the LGN has about twice the number of neurons compared to V1. However, experiments …
###Reviewer #2:
The authors introduce a computational model of the interplay between excitatory and inhibitory plasticity during development in V1. The analysis of the work is interesting; however, several assumptions have to be checked and a multitude of additional analyses is required to validate the conclusions.
Major Comments:
The model describes the dynamics during the development of V1. However, during development there are several phases, each having its specific properties and dynamics. For instance, van Versendaal and Levelt 2016 discuss that especially inhibition could have a critical and phase-specific role. Please discuss in more detail the relation of the model to the developmental periods or rather which period you model.
In the model, the LGN has about twice the number of neurons compared to V1. However, experiments estimate that V1 has 40 times more neurons than LGN yielding a different type of projection. Please test the dynamics for a significantly larger V1. Furthermore, please test the dynamics resulting from a sparse connectivity between areas, as all-to-all connectivity is a very strong assumption.
The authors neglect recurrent excitatory-excitatory connections. Please show at least the influence of non-adaptive recurrent excitatory connections on the results.
In the model, the role of inhibition is mainly to constrain the neuronal activities, which can also be done by other homeostatic plasticity mechanisms. Would intrinsic plasticity also be sufficient? Also the role of homeostatic synaptic plasticity for V1 development has already been shown in other computational studies (e.g., Stevens et al., 2013; J. Neurosci.). Please discuss.
In general, EI2/1 seems to be more efficient than EI3/1. What is the lower limit? Is an EI1/1 system even better? In addition, the reduction of redundancy could imply that the system becomes less robust against noise. Please test for different noise levels/sources and whether noise implies a lower bound.
The authors discuss on Page 18 that the learning rates of the involved plasticity processes are important. However, they do not show any data. Overall, the parameter-dependency of the model remains unclear. Especially given that the parameters of inhibitory plasticity are not based on experimental data, these have to be investigated in more detail.
The authors say that the receptive fields in the model are stable. Please show any data supporting this claim. Under which condition are the receptive fields stable?
Is the model leading to any experimentally verifiable predictions?
-
###Reviewer #1:
This manuscript details a modeling study used to understand how inhibitory plasticity shapes the emergence and structure of receptive fields in visual cortical networks. The work seems well-carried-out and the writing is clear.
Major concerns:
It needs to be made more clear in the manuscript how these results extend on what has been shown previously on the emergence of V1-like RF's in cortical networks. The new insight here is not apparent in the framing of the introduction. A somewhat more detailed answer to the question "How surprised should one be by these results?" particularly about the emergent gain adaptation, would be useful.
It would be very good to see more comparisons between fixed inhibition and inhibitory plasticity in this work, especially since this is advertised in the title and abstract as the main thrust …
###Reviewer #1:
This manuscript details a modeling study used to understand how inhibitory plasticity shapes the emergence and structure of receptive fields in visual cortical networks. The work seems well-carried-out and the writing is clear.
Major concerns:
It needs to be made more clear in the manuscript how these results extend on what has been shown previously on the emergence of V1-like RF's in cortical networks. The new insight here is not apparent in the framing of the introduction. A somewhat more detailed answer to the question "How surprised should one be by these results?" particularly about the emergent gain adaptation, would be useful.
It would be very good to see more comparisons between fixed inhibition and inhibitory plasticity in this work, especially since this is advertised in the title and abstract as the main thrust of the work. In the current draft, this is addressed only in Figure 9 but should play a more major role throughout the draft, to strengthen and emphasize the novelty of the work.
Some amount of theoretical work to complement the simulations would strengthen the paper greatly.
Comparisons to other plasticity models, to show what exactly is necessary for replicating the effects here seems very important, but under-explored.
When speaking about metabolic costs of computation, it seems important to also discuss the size of the network and the maintenance of synapses, not just the average firing rate per cell. Some discussion of this should be included, or some of the claims in the intro/abstract should be softened.
-
##Preprint Review
This preprint was reviewed using eLife’s Preprint Review service, which provides public peer reviews of manuscripts posted on bioRxiv for the benefit of the authors, readers, potential readers, and others interested in our assessment of the work. This review applies only to version 2 of the manuscript.
###Summary:
This manuscript details a modeling study used to understand how inhibitory plasticity shapes the emergence and structure of receptive fields in visual cortical networks. The work seems well-carried-out and the writing is clear. The authors implemented a recurrent network with excitatory plasticity and inhibitory plasticity at all connections - both feedforward and recurrent. The results reveal that a model with inhibitory plasticity exhibits more diverse receptive fields (covering the different orientation …
##Preprint Review
This preprint was reviewed using eLife’s Preprint Review service, which provides public peer reviews of manuscripts posted on bioRxiv for the benefit of the authors, readers, potential readers, and others interested in our assessment of the work. This review applies only to version 2 of the manuscript.
###Summary:
This manuscript details a modeling study used to understand how inhibitory plasticity shapes the emergence and structure of receptive fields in visual cortical networks. The work seems well-carried-out and the writing is clear. The authors implemented a recurrent network with excitatory plasticity and inhibitory plasticity at all connections - both feedforward and recurrent. The results reveal that a model with inhibitory plasticity exhibits more diverse receptive fields (covering the different orientation preferences more uniformly) compared to a model without any inhibition but with plastic excitatory synapses. Synaptic connectivity reflects tuning similarity, and inhibition aids in decorrelation. In this model, inhibition sharpens tuning curves, helps to develop contrast invariance, and promotes sparseness. Finally, the manuscript shows that the plastic model has a lower reconstruction error compared to a model without inhibition at all.
The reviewers found the results presented to be clear. The reviewers also thought that some new analyses should be done to shore up the results, and that writing revisions could be implemented to improve the flow of ideas for the reader.
-
