Explaining V1 Properties with a Biologically Constrained Deep Learning Architecture

Summary

In this study the authors propose the incorporation of mechanistic biologically inspired filtering and normalization components in deep convolutional networks (DCNs) with the goal of increased alignment of model responses to V1 neural responses and tuning properties. The authors add center-surround receptive fields, local receptive fields, tuned divisive inhibition and cortical magnification to DCNs that are trained with downscaled ImageNet-64x64. The authors perform an extensive ablation of the components above to show their relative importance for alignment with V1 neural responses. The proposed best-performing model produces quite significant improvements on Brainscore's V1 model alignment. The authors also touch upon whether the above models with high V1 alignment are more robust to perceptual distortions.

Strengths

+ The authors propose a unique combination of biologically inspired components that have been shown to exist in primate early visual mechanisms. These components are also explained in good depth and clarity for readers who may not be familiar with the specific computations. To the best of my knowledge, even though each of these components by themselves are not novel, the unique combination explored here appears to be novel and not explored before from the perspective of bettering alignment to V1 properties. + The proposed model is significantly outperforming the previous SOTA on explaining neural activity and tuning properties from the Brain-score dataset. The authors have run multiple simulations with different random seeds and add more credibility to their observed findings.

Weaknesses

- As the authors have pointed out, there is a significant drop in accuracy with respect to image classification accuracy. I believe the authors must try to address, from their perspective, why this drop occurs. It is common in this area that models that are trained to better represent biological neural activity tend to suffer from poor classification accuracy. Addressing this issue will be quite a strong contribution. - The scope of this work is quite limited; I am unsure if the produced improvement on alignment to V1 neural responses and tuning properties is sufficient as the only major contribution in this work. - It may help if the authors could please add some intuition about why each of the explored components help in improving neural predictivity; it is also important to find out whether these components are useful to improve model-brain alignment regardless of the underlying architecture.

Questions

Please see my review in the Weaknesses section for questions and suggestions.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, the authors have adequately addressed limitations in their work.

Summary

The paper considers deep networks as a model of the visual stream, specifically V1. The authors systematically study the impact of various biological additions to deep networks on alignment of deep nets' representations with V1 recordings.

Strengths

The paper considers several features of the early visual stream that can be added to deep networks, and tests the influence of those features on image classification performance and V1 alignment. While some individual features have been considered before, the approach and especially ablation studies here are novel and, in my opinion, interesting to the community. The final result is interesting too: combining all 4 architectural features alone resulted in the best V1 alignment (0.605 vs. 0.594 of the top1 V1 model www.brain-score.org/model/vision/623). Adding adversarial features improved it to 0.629, which seems very significant -- the median V1 score at www.brain-score.org/ is less than 0.5.

Weaknesses

The results in Tab. 3 suggest that V1-like features significantly hurt ImageNet performance -- the best V1 model is 16% less accurate than the best ImageNet model. This is mostly due to adversarial training. I think the authors should discuss why it has such an effect. Two important ablation studies are missing: 1. Adversarial training only, since it has a big effect on both V1 alignment and ImageNet performance. 2. Untrained networks with all/some biological features. All discussed features change the distribution of neural responses even in untrained networks, so it might be that V1 improvements come from that distribution change alone, not from training with those features. I also suggest the authors include, at least in the appendix, Brain-Scores for other areas (V2, V4, IT) and behavioural data. This is the standard way to evaluate models on Brain-Score, so having all results would make comparisons to other models easier.

Questions

Is it possible to provide a baseline for the "best possible" V1 score by comparing Brain-Score neural data to itself (e.g. with K-fold cross-validation)? I don't think it was done in the original Brain-Score papers, so it's definitely not a hard requirement here. But it would be a great addition. ### Minor issues > [Line 20] Advances in neuroscience have long been proposed as essential to realizing the next generation of artifical intelligence (AI). Is that true? I'm not quite sure… also misspelled artificial > [22] (e.g, convolutional neural networks and mechanisms of attention) owe their origins to biological intelligence Conv nets need a citation; attention too, and I’m not sure if attention mechanism in transformers were even inspired by biology on the implementation level (see https://www.frontiersin.org/articles/10.3389/fncom.2020.00029/full which says "While the spirit of attention in machine learning is certainly inspired by psychology, its implementations do not always track with what is known about biological attention, as will be noted below.") Overall: \cite doesn’t generate links to bibliography? > [45] In specific I don’t think this phrase is commonly used. “Specifically” or “in particular” would read better. > [151] these DoG convolutions were only applied to a fraction of the input feature map Why? Fig. 1D can benefit from a more detailed explanation. I want to say the original image is on the right but then the transformation doesn’t preserve retinotopy. > [209] alternate Alternative Color-coding Tab. 3 would be great! The code is in the supplementary, so the authors should indicate it in the main text (and perhaps add it to github afterwards) ### Rebuttal acknowledgement I have read the rebuttal and responded to the authors. I think it addressed all (minor) concerns that I had, and I still think 7 (accept) is an appropriate score.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The limitations and impacts are addressed.

Summary

The authors incorporated four well-known architectural components of V1 into an earlier layer of the CNN, resulting in a reduction in task performance but an improved alignment with V1 neurons' behaviors. Their study demonstrated that cortical magnification led to the most significant enhancement in alignment, as observed in the overall property and predictability of V1 in the Brain-Score benchmark test. Tuned normalization also improved alignment in certain V1 properties, while the contribution of center-surround mechanisms appeared to be minimal or data-dependent. These improvements generally ranged between 1% and 2%.

Strengths

The motivation and hypotheses of the study are reasonable, and the exploration is conducted in a systematic and logical manner. The finding that cortical magnification provides some improvement in alignment is interesting.

Weaknesses

The introduction of brain architectural components was expected to enhance the alignment of the model with V1 data in the Brain-Score test, so the results are not particularly surprising. While the paper contains some systematic and well-done experiments, it does not provide an explanation as to why certain architectural components would have specific effects. As a result, the main contribution is simply showing incorporating more information relevant to the data would improve performance in explaining the data, at the expense of the model's task performance. Although the work may hold value, its contribution might not reach the level typically expected in a NeurIPS paper.

Questions

It would be helpful to discuss why and how certain architectural components would produce greater alignment while others do not. Is there any theoretical and conceptual framework that can help us to make sense of the results?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations on performance drop has been discussed. The work could have implications on neuroscience.

Summary

This paper incorporates a wide range of biologically inspired components into the initial stages of a ResNet model to see if these result in improved alignment with properties of V1 neurons. Specifically, the authors incorporate architectural components for Center-Surround, Local Receptive Fields, Divisive Normalization, Tuned Normalization, and Cortical Magnification. They find that some of these components are complementary when explaining different properties of V1. Adversarial training further improves the match to V1 responses.

Strengths

This paper provides a through discussion about many hypothesized computations in V1, and incorporates these computations into deep neural networks as modules with learnable components. The individual contributions of these components are systematically evaluated in terms of how well they capture properties of V1 and predict V1 responses to stimuli. Building this type of biologically inspired model with many V1 components is novel and the discussion of incorporating biological components into AI systems is currently of high interest for the NeurIPS audience (both on the CS and neuroscience communities).

Weaknesses

* The idea of including biological components into neural networks is not novel, and the paper lacks full discussion of other attempts along these lines. Perhaps most relevant here is VOneNet (Dapello et al. 2020) which incorporates properties of V1 into a convolutional neural network (this paper is cited but not in the context of modeling V1). Although the exact architectural components added are different here (and designed for specific response properties), the paper would benefit from an explicit discussion about what sets this work apart from what was previously done. * Re clarity: As the paper is currently written, I found difficult to follow what is previous work and what is built into the model. In some parts of the “Background” it is mentioned that certain things are adopted in the previous work, but other parts are not discussed as being incorporated until the Methods. It might help the reader if these sections were restructured so that it is clear how the previous work builds into the proposed architecture. * The paper claims to “introduce architectural components based on neuroscience foundations” however, from what I can tell, all of the included architecture components have been previously proposed. The novelty of this paper is including all of them within one model and analyzing them systematically. * Due to the number of models and response properties that are tested, and the somewhat small differences in many cases, it is difficult to interpret which of the brain-score results are significant, although I do appreciate that the authors report mean and standard deviation across 3 trials of training and evaluation. Perhaps just changing the colors to make it clear which differences are or are not significant (correcting for multiple comparisons) would help? But generally, the changes in Brain-Score seems relatively small as all models are still far from the noise ceiling, making some of the claims of the paper seem too grandiose given the reported results. * The robustness to corrupted images experiment also seems like the differences from the baseline model may be within the noise after correcting for multiple comparisons, and the overall changes are very small.

Questions

a) It would be helpful to include a model with *just* adversarial training and none of the biological components as an additional baseline model in Table 3. b) I’m a bit confused about the greedy backwards elimination approach. Was the architectural component that “reduced overall V1 alignment” selected specifically at that given stage of the architecture (ie was a model trained eliminating each component separately, and then just one was chosen?) Or was this elimination order based on the single-architectural component experiment? c) On line 199-200 it states that the cortical magnification is applied immediately before the ResNet50 layer 1. Does this ResNet still have the first convolutional layer that proceeds all of the residual blocks, such that the cortical magnification is acting on the output of the convolutional layer/pooling? This detail should be clarified. d) Was the decision to downsample the ImageNet images to 64x64 based on biology or computational constraints? Is it possible to train on more standard image sizes? This would possibly help the accuracy of the models. e) Related to the above – is the baseline ResNet50 presented trained on these 64x64 images as well? f) What networks are being referred to on lines 327-330? Are these the networks with adversarial training? (If so, this claim maybe should be toned down, as the margins are not “large” and still far off from the noise ceiling). Minor: * Lines 25-28: It would be helpful if specific “neuroscientific” models were listed here. I think the idea is that the models referred to in this section are hand-designed based on observed neural properties, rather than being optimized in some other manner? * Line 32: “Through typical task-driven training alone” – this is not quite true, as a (typically linear) readout must be trained to map the activations of the neural network onto the responses of neurons. * Lines 36: It would improve the paper if citations showing that CNNs are not achieving properties of the visual system were included here. * Lines 68-73 are vague and uninformative to the context of the presented work. It would help to be more specific about failures of the models and previous work. * Line 325: Include a reference to the supplementary section documenting the training details for the adversarially trained network.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

The authors discuss the limitations in the discussion.

Response to rebuttal

Thank you for the response! Overall, I think it addressed all (minor) concerns that I had, and I still think 7 (accept) is an appropriate score. > We demonstrate the efficacy of this strategy with a naive proof of concept in which classification is performed by averaging predictions from the five static crops of each image Great! I believe this technique, test-time augmentation, is not uncommon in deep learning, and makes sense for models of the visual stream. I also think the new results with untrained networks, which achieved much lower V1 alignment than the trained ones, strengthen the results -- they suggest that all added components don’t just match superficial features of V1 processing, but lead to better (in terms of alignment to V1) training. Other reviewers have noted the limited scope of this paper, but I don’t completely agree. I think the contribution of this paper is significant enough for the task of building better models of the visual stream.

Response to author rebuttal

Thank you for the response. From this, it is more clear that the benefit of this work comes from integrating all of the different V1 components into the same model, showing that this results in better alignment with V1 responses. If this is the case, I think that should be made more explicit throughout the paper (ie in the listed bullet points on line 46). I also agree with Reviewer yZMG about the Untrained models being a nice addition. After the author responses I still find this work borderline. I appreciate that amount of effort it takes to integrate all of these components into a single model (and that itself may be useful for the field as a new baseline, as the authors discuss). But as currently presented, I'm not sure if that is "new" enough or if the insights gained are deep enough for a typical NeurIPS paper. Finally, this is a bit of my personal preference in terms of wording (and thus is not influencing my score), but as neuroscientists, I wonder if we want to fall into the trap of publishing paper after paper highlighting the achievements of "SOTA" on a particular benchmark when the improvements are around +2-3%? To me, it seems especially problematic given the limited size of the current datasets. This is why I was highlighting the concerns about multiple-comparisons above, and suggestions to tone down the language about things like "unprecedented explanation" of V1.

Thank you for the author rebuttal

I thank the authors for responding to the reviewers' concerns in the rebuttal. I appreciate the authors effort to successfully reduce the drop in ImageNet accuracy using static crops of the images. I do agree with the reviewers that the proposed work will be valuable to neuroscientists as a model of V1 processing, I am slightly improving my score from my pre-rebuttal evaluation of the paper.

Thanks for the responses

Thank you for the responses. It is indeed interesting to show that "it was only when these components were combined that we observed drastically improved explanations of V1, suggesting their complimentary contribution." Somehow I did not get this on my earlier reading. Thank you also for the additional experiments which do provide some insights. I am willing to upgrade my score, but my overall sentiment is very much aligned with that of Reviewer bbGj. Incidentally, it would also be interesting to compare this transfer-learning + neural constraint model with purely data-driven model, e.g. https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1006897 However, I suppose that data set won't have contextual modulation effects, and thus can't serve as a real neuron-in-silico. Perhaps some Allen Institute dataset on mice could work.