NeurIPS 2020

A simple normative network approximates local non-Hebbian learning in the cortex


Review 1

Summary and Contributions: The authors derive online versions of RRMSE and CCA, and map it to pyramidal neurons with distal/proximal inputs, inhibition and Ca^{2+} plateau potentials. They show empirical results that these local online rules work well.

Strengths: The authors derive simple models of the online variants of RRMSE and CCA, which is new to my knowledge. The use of such local learning rules might lead to efficient implementations, say on neuromorphic hardware.

Weaknesses: The empirical evaluation is one of the weakest aspects of the paper. The fact that this is done on only one, seemingly arbitrarily chosen, dataset diminishes the significance of the results. I would have liked to see evaluation on more standard datasets. There are some aspects of the biological mapping that may not be biologically plausible: - Only linear model are considered. In biology, pyramidal cells are known to have many non-linear effects. - Inhibition is considered with tied input and output weights. Need more details on which type of inhibitory neurons ave this property. - Distal plasticity ignores known data from (Milstein et al. 2020; Magee and Grienberger 2020 [1]) the first of which is cited in the paper for other reasons. - The authors don't consider dendritic spikes and calcium plateau potentials as events. - V_y updates use activity of soma and input to distal compartments, which is quite unusual for biological rules. - Ca^{2+} plateau potential are generally know to affect distal dendrites, which is not considered here. - They don't consider binary or spiking data which could be more biologically plausible. An analysis of what happens when moving from full dataset to single samples would have been useful. The authors consider algorithms that are used for dimensionality reduction (as mentioned in their discussion), but need to specify which pyramidal cells in the brain are known to have this property. A discussion of how representations learnt using RRMSE or CCA can be used by further processing layers can strenghten the paper. [1] Magee, J.C., and Grienberger, C. (2020). Synaptic Plasticity Forms and Functions. Annual Review of Neuroscience 43.

Correctness: Some of the decisions taken when mapping to biology ignores or stretches known data. The empirical evaluation is done on only one, seemingly arbitrarily chosen, dataset.

Clarity: The paper is generally clearly written, and the authors motivate the problem quite nicely.

Relation to Prior Work: (Milstein et al. 2020; Magee and Grienberger 2020) discuss potential plasticity mechanisms where Ca^{2+} plays a role, which is not taken into consideration. (See list of biologically implausible aspects of the model.) Other work on role of pyramidal neurons and plasticity there [2,3,4] could be discussed. [2] Larkum, M. (2013). A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex. Trends in Neurosciences 36, 141–151. [3] Kampa, B.M., Letzkus, J.J., and Stuart, G.J. (2007). Dendritic mechanisms controlling spike-timing-dependent synaptic plasticity. Trends in Neurosciences 30, 456–463. [4] Gidon, A., Zolnik, T.A., Fidzinski, P., Bolduan, F., Papoutsi, A., Poirazi, P., Holtkamp, M., Vida, I., and Larkum, M.E. (2020). Dendritic action potentials and computation in human layer 2/3 cortical neurons. Science 367, 83–87.

Reproducibility: Yes

Additional Feedback: Update post-author-response: I am satisfied with the new empirical results. The author also promise to make it very clear which biological aspects are supported by currently known data and which is speculation. I am revising my score to accept. Minor: On l.121, the authors probably mean eqn. (3) not (18) Is the MediaMill dataset conducive for linear models? What happens on a dataset that is not, with these learning rules?


Review 2

Summary and Contributions: RRR and CCA are two classical statistical techniques with well-known solutions. In this paper, the authors design new online learning algorithms for RRR and CCA, such that learning works under biological constraints, or, specifically, such that all learning is local. The results are given a biological interpretation and are demonstrated in numerical experiments.

Strengths: Overall, I found the paper interesting. In principle, RRR (and CCA) could be implemented by two-layer linear networks, and then learnt via back-propagation, for which, by now, several candidates for local learning rules exist. However, the authors here choose a different path, which I think is interesting in its own right. Essentially, they eliminate the middle layer, design a novel online algorithm and then go to great length to give their algorithm a biophysical interpretation. I think this new way of thinking about what essentially is a two-layer network provides an interesting perspective on a classical learning problem.

Weaknesses: The biological interpretation is maybe the weakest part of the paper, as it is quite speculative, but the authors do not clearly state that. Basically, I find the online algorithm intriguing, and I find that the biological interpretation consists of mixture of interesting ideas and some claims that are plainly wrong (see below).

Correctness: The neural circuit implementation suffers from the fact that it's a loose interpretation of the terms in equations (9)-(11). These equations are, at most, those of a rate network, and there is no established mathematical framework to consider compartments, plateau-potentials, etc. in rate networks. Therefore, many aspects of the biological circuit implementation hang in the air, and seem speculative at best. Some also seem simply wrong. (a) For instance, the output of the pyramidal cell is assumed to be z=Vx. That pre-supposes that inputs from the apical dendrite (a-Qn) either cancel out to zero, or never make it into the soma. Either seems wrong, especially if those inputs are supposed to give rise to plateau potentials which will almost certainly cause rapid firing of the output. (b) The term a-Qn is supposed to be a 'calcium plateau potential traveling down the apical shaft.' However, Calcium plateau potentials are really like slow calcium spikes, meaning that, to first order, they are stereotypical events. They are not good candidates to carry graded signals. (c) Also, you use pyramidal cells and interneurons, but they are not excitatory or inhibitory neurons (as both cell types can excite or inhibit their downstream partners). Please clarify. More generally, I dont mind (even wild) speculation about matches to biology, but it should be made crystal clear what's speculation, what's a hypothesis, and what's fact. The paper does not generally succeed in doing so.

Clarity: The technical parts are generally well written. The one section I struggled with was the link to backprop (Figure 2), which is quite intriguing, but I found the text and logic hard to follow. I think it would be of interest to the field if the authors could expand on these ideas (as many people are interested in biologically plausible backprop)

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback:


Review 3

Summary and Contributions: This paper finds a local, online algorithm for performing reduced rank regression and canonical correlation analysis, and also any interpolation of the two. Update post-author-response: thanks for your reply, but linearity was not a big issue for me.

Strengths: The approach is elegant, and the circuit matches what's seen in hippocampus and neocortex -- including plateau potentials.

Weaknesses: No major weaknesses. A very minor weakness is that they're essentially doing linear dimensionality reduction, but it's likely that's a computation that the brain does.

Correctness: I think so.

Clarity: Yes. A bit dense, but that's because of the subject matter.

Relation to Prior Work: I'm not an expert, so I can't really say.

Reproducibility: Yes

Additional Feedback: None.


Review 4

Summary and Contributions: ***** UPDATE ***** I appreciate the new empirical results and am increasing my score by 1 point as a result. My remaining hesitations center on whether the paper effectively balances (1) model simplicity, (2) ability to account for experimental findings and (3) a computationally powerful algorithm. I was hoping for more on (1) and (2) in part because RRR is a relatively computationally weak/simple algorithm - maybe on par with PCA, which can be accomplished with a much simpler model like Oja's Rule. There have also already been published a number of multi-compartmental neuron models that implement interesting algorithms while explaining some subset of experimental findings (and I suspect that many more are possible), and it's not clear to me that this model is more parsimonious or explanatory than existing ones. If accepted, I would be happy to see some of the above briefly addressed in the paper. ***** The authors develop offline and online learning rules defined to achieve a Reduced Rank Regression objective. By mapping the online rule onto various intra-neuron signals in a neural model with segregated proximal and distal compartments, they develop the rule into a model of local cortical plasticity driven by supervisory signals. They further present the experimental evidence in support of the neural dynamics predicted by the online rule, and compare the performance of the offline and online rules to several baselines on the MediaMill benchmark dataset.

Strengths: The paper is well-written and theoretically well-developed. Since the authors start with a family of well-defined and well-known objectives in RRR and work backwards to the learning rules, the results are less dependent on simulations. Finally, the paper does an admirable job of reviewing the experimental evidence and discussing potential issues with the model in the final section.

Weaknesses: The normative approach is appealing, but the model lacks any form of temporal dynamics which is limiting as a biologically plausible model of neural dynamics or plasticity. The linear approach is also limited both as a description of neurons and in computational capacity. These limitations are challenging in comparison to existing models of learning dynamics built on segregated neuron models with richer dynamics (e.g. Guergiuev as cited, or Urbanczik and Senn 2014 below). I would like to see a more convincing argument in the Introduction for the significance of this new model over the existing models. While the authors nicely show how the learning rule can be interpreted in terms of prediction error "at optimum", I would like to know whether the interpretation approximately holds more generally (e.g. during learning); potentially in the supplemental.

Correctness: The claims, method and methodology appear to be correct.

Clarity: Generally the paper is well-written. A few more sentences describing the baseline algorithms in Section 6 would be appreciated, if it can fit.

Relation to Prior Work: As mentioned above, I think that the paper would benefit from a clearer argument for the value of this model over similar existing learning rules. It's not clear to me that the existing models "optimize only the synaptic weights while postulating the architecture and activity dynamics" more than this work. To the list of existing research, I would add Urbanczik and Senn, "Learning by the Dendritic Prediction of Somatic Spiking", which similarly develops a compartmental learning rule with teaching signals.

Reproducibility: Yes

Additional Feedback: