NeurIPS 2020

Lamina-specific neuronal properties promote robust, stable signal propagation in feedforward networks

Review 1

Summary and Contributions: This paper redid the classic Ad Aertsen paper in which they asked whether synchronous volleys of spikes could propagate through multiple layers of a network. Unlike the original work, the goal here was to preserve the temporal width and firing rate of the synchronous volleys. By alternating neuron types in adjacent layers, they largely succeeded.

Strengths: The main strength is that they got it to work.

Weaknesses: The weakness is that it's not clear why this is important. If the brain wants to send information, it will probably use axons; if multiple layers are involved, it's doing some sort of computation. Update post-author-response: In their reply, the authors pointed out that their method would be used for non-connected areas (since the brain presumbably can't route new axons). I'm skeptical: it would have to route information through connected areas, using axons that are already used for something else. That seems insanely hard. Also, I'm worried about robustness. By using two neuron types, when the width (sigma) and amplitude (a) of a pulse go through two layers, they are transformed as [a_(l+2), sigma_(l+1)] = f(g(a_l, sigma_l)). where both f and g are vector functions with two outputs. In their simulations, f(g(a, sigma)) \approx [a, sigma]. More accurately, there seemed to be several fixed points in the [a, sigma] plane. This is non-generic behavior, and one wonders if there was fine tuning involved. Update post-author-response: the authors said no fine tuning, and pointed to Fig. 3E. To be convinced, I would need to see a parameter sweep.

Correctness: I think so.

Clarity: Yes.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: None.

Review 2

Summary and Contributions: Feedforward networks (FFNs) with heterogeneous laminar-specific neuronal properties are shown to stabilize and enhance information transmission compared to homogeneous FFNs. Simulations results captured key experimental properties of the multilayered circuitry of the Drosophila olfactory system. The interplay between integrator and differentiator neuron types across network layers compensates distortions introduced by each other and provides robust spiking information transfer. Having read the author's reply and the other comments, I stand by my original assessment.

Strengths: The work provides simulation results that extensively describes how a particular heterogeneous multilayered circuitry could improve accuracy and speed of information transfer in comparison with the equivalent homogeneous circuitry. It raises the point that heterogeneity in properties of neuron types is a critical factor for robust information transfer, in particular when considering multilayered feedforward networks.

Weaknesses: The main conclusions of this work are particularly well suited to the experimental findings of the specific multilayered circuitry of the Drosophila olfactory system. However, it is not clear how general these conclusions are and if they could be extrapolated to other brain areas and other widely-studied neural architectures. For example, inhibitory neurons play an essential role in cortical processing but this work partially addresses their influence including only some results with some sort of unspecified feedforward inhibition in Appendix Fig. A

Correctness: Methodology is well-defined, proven, and thoroughly documented. Claims are correct based on the proposed methodology.

Clarity: The paper is well written.

Relation to Prior Work: It is clearly discussed how this work differs from previous modelling studies of FFNs. However, it would be interesting to note that although previous works did not explicitly model heterogeneity across layers, they did introduce heterogeneity into the network choosing parameters of neurons (e.g., Vth or C) from probabilistic distributions (see for example the work by Kumar and colleagues, 2008, the Journal of Neuroscience).

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: This paper uses the Drosophila’s olfactory system (ORN to PN to LHN) as an example to demonstrate how the heterogeneity in the intrinsic parameters of single neurons could benefit for information transmission in a feedforward network, which is the main conceptual contribution of this work. Then they extend the idea of heterogeneous neurons to multiple layers of feedforward networks.

Strengths: The authors did numerical simulations of the network model to support their claims. In particular, the network parameters are consistent with real neural circuit, in that they were fitted from experimental data. The simulation results are solid and clearly demonstrate with a repetition of integrator-differentiator networks, the firing rate profile of input spikes could be reliably transmitted through multiple layers (Fig. 3).

Weaknesses: Although I believe the intrinsic difference of neurons could benefit for information transmission, I have some conceptual questions. I think properly answer these questions in the Discussion or briefly mention some of them in author feedback could improve the impact of this work in general. 1. Whether a network is an integrator of a differentiator is highly determined by the value of \beta. Is it possible with an intermediate value of \beta, the network’s output is proportional to the input, i.e., the network simply relay the input but neither differentiating or integrating. In this case, we probably only need one layer to transmit the input without the cascade of an integrator and a differentiator. Update after rebuttal: I think in principle there might be a particular \beta to achieve same transmission with integrator-differentiator circuit, but some fine-tuned mechanisms are needed. And hence the heterogeneity would be a more robust mechanism to achieve reliable transmission. 2. If we reverse the order of differentiator and integrator, would the input be reliably transmitted as well? Is there some considerations for Drosophila’s olfactory system has a structure of an integrator followed by a differentiator? Update after rebuttal: I agree author’s statement about this and I hope the author could briefly discuss the order of integrator and differentiator in a revised manuscript.

Correctness: My own concern is about the conclusions based on Fig. 4 (correct me if I was wrong). How does Fig. 4A could lead to the conclusion that the proposed network model could adopt both spike rate and spike timing coding (line 228)? Update after rebuttal: I am still not clear about the justification of this. Please provide more elaborate claims in a revised manuscript.

Clarity: This paper is well written, and structure-wise.

Relation to Prior Work: The comparison with previous work was mentioned in Discussion. The authors mentioned earlier models need fine-tuned to achieve reliable transmission, while in current study the stable transmission was achieved by heterogeneity on intrinsic parameters of single neurons. One thing is not very clear is that have previous network models considered the heterogeneity on the intrinsic parameters of neurons? If not, I hope the author could clearly mention this point in a revised version to better emphasize the contribution of this work.

Reproducibility: Yes

Additional Feedback: 1. I am quite confused about the dynamics (Eq. 1), although I believe the authors used a correct dynamical equation in their simulation. The thing puzzled me a lot is that there is no interaction at all be 2nd row and and 1st row in Eq. 1. Moreover, the 2nd row says the variable z could be either m or w, and why the authors write a separate expression of w in the left most on the 2nd row? 2. Fig. 2C and Lines 144-146: I am a little concern about whether the comparison between the heterogeneous network and homogeneous network is fair. In computation, I am wondering whether the \beta could be adjusted into an intermediate value and then the cascade of two homogeneous networks is the same as a homogenous network? 3. I will be very glad to see if authors could provide a figure demonstrating how the network varies from an integrator to a differentiator when the value of \beta changes. 4. I am also curious to know how sensitive the reliable transmission depends on the values of \beta in two consecutive layers.