NeurIPS 2020

Neuron Shapley: Discovering the Responsible Neurons


Review 1

Summary and Contributions: The paper proposes to use Shapley Values to find a small number of hidden neurons that are "most responsible" with respect to a tunable objective, i.e., the overall performance of the network, the recall of the network for one specific class, its bias towards certain data distributions or its sensitivity to adversarial perturbations. The authors show that Shapley Values are more effective at determining the few most relevant neurons compared to Neuron Conductance in a number of experiments in the image domain. Moreover, in order to approximate Shapley Values in a reasonable time on large networks, the paper proposes a novel algorithm that combines Monte-Carlo sampling with a multi-armed bandit.

Strengths: The choice of employing Shapley Values to compute the importance of hidden neurons is well theoretically justified and inherits from related literature in the XAI domain. While the idea of estimating the importance of hidden neurons is not new (see, for example, the cited paper of Neuron Conductance), the idea of using multi-armed bandit for determining the few playes with highest attribution is novel and relevant to the community. Although the method is developed with the idea of determining which are the top-k neurons (rather than estimating the Shapley value for all neurons), I find it interesting that the rank correlation with monte-carlo sampling is so high. I imagine that the same method can also be used to find the few most relevant input features for a given prediction. The paper also shows some interesting phenomena about convolutional neural networks, mainly the fact that there exist a few filters that are responsible for most of the unfair behavior of a classifier or for its sensitivity to adversarial perturbations. Although each phenomenon was not investigated in details, these are still interesting observations in the context of fairness, transparency and security, which are important topics at the moment.

Weaknesses: The idea of applying Shapley values for the understanding of deep neural networks is not new. Several works, such as Lundberg et al., 2017, have already discussed the theoretical motivation for using Shapley values as an attribution method to rank the importance of the input features. Lundberg et al., 2017 also proposed approximations like KernelSHAP and DeepSHAP, which are not compared to TMAB-Shapley. Besides this line of works, the idea of using Shapley values to rank the internal neurons has been proposed by the Stier et al., 2018 (cited) and Florin Leon, 2014 (not cited) in the context of pruning. Finally, Ancona et al., 2019 (not cited) proposed an approximation technique for Shapley values tailored for deep neural networks. How does their method compare with TMAB-Shapley in terms of accuracy and performance? At least a theoretical comparison with Ancona et al. and Lundberg et al. would be needed to justify the authors' claim "this algorithm enables us, for the first time, to compute Neuron Shapley values in large-scale state-of-the-art convolutional networks". I also believe that the experimental section is somewhat superficial. While providing several experiments, the paper does not go in-depth in any of them. This leaves several open questions: - [Neuron Shapley identifies a small number of critical neuron] While it is interesting to see that the network mostly relies on a few very important neurons, what are the practical implications of that? Does it mean that it is possible to use this technique the other way around, and find the least important neurons that can be pruned? This would be a more relevant use case I believe, but it would also require a comparison with other structured pruning techniques. - [Class specific critical neurons] the baselines of comparison are rather weak as one is random and two do not consider how the information is processed after the activation (l2 norm of the weights and l2 norm of the response). I would expect Neural Conductance to be a comparison benchmark for each experiment, in this case, to be shown in Fig 2a. - [Discovering unfair filters] Does the inspection of the "unfair" filters and their activating samples from the training set shed any light on what these filters are capturing? - [Identifying filters vulnerable to adversaries] I did not quite understand the sentence "We note that while the modified network is robust to the original adversary, it is still vulnerable to a new adversary specifically designed to attack the modified network. This requires a white-box adversary who knows exactly which few neurons are zeroed out." Does it mean that the accuracy was tested using the same adversarial images pre-computed for the original network? If it is the case, and if the new network can still be fooled by running PGD on it, how can the method "offer a fast mechanism to repair the model"? The authors would need to show that at least for some attacks (either white or black box) the robustness is improved when the new model is attacked from scratch. The comparison with Neuron Conductance, which is the only strong baseline, seems to be unfair as 25000 images are used to compute Neuron Shapley and only 100 to compute conductance (according to Appendix C). Why this is the case? This means that conductance is computed without providing not even 1 example per class. The reported computational time is also misleading, as conductance can be computed much faster (https://arxiv.org/pdf/1807.09946.pdf, a possible implementation https://captum.ai/docs/algorithms). Since the authors are using a quite unique setup (a large CPU cluster), I would suggest reporting the running times in terms on the number of network evaluations needed (either forward or backward). - [1] Ancona et al., 2019, Explaining Deep Neural Networks with a Polynomial Time Algorithm for Shapley Values Approximation, ICML 2019 - [2] Florin Leon, Optimizing Neural Network Topology Using Shapley Value, 2014, ICSTCC

Correctness: Some claims require clarification: - lines 66-74: 1) what are the "fair credit allocation properties"? According to the original paper, Neuron Conductance also satisfies Zero contribution, Symmetry and Additivity so what are the additional properties that Neuron Shaply satisfies? This section does not mention that gradient-based methods are usually much faster than sampling-based methods. - line 134: while the definition considers all possible coalitions, the marginal contribution of each feature is averaged over all coalitions thus Shapley values is still a linearization of the model and some cross-dependencies might be hidden in the average. For example, while it is guaranteed that a null player will be assigned zero attribution, the other way around is not true: if the Shapley value is 0, the neuron might still have a significant role. Also, the paper says "is one of the few methods that take such interactions into account". What are the other methods? - line 210: there is a vast literature on pruning neurons or filters (structured pruning) like [1-4] - as mentioned before, it is not clear to the reviewer how the results on adversarial examples are computed. [1] HE, Yihui ; ZHANG, Xiangyu ; SUN, Jian: Channel Pruning for Accelerating Very Deep Neural Networks. In: The IEEE International Conference on Computer Vision (ICCV), Oct 2017 [2] HU, Hengyuan ; PENG, Rui ; TAI, Yu-Wing ; TANG, Chi-Keung: Network trimming: A data-driven neuron pruning approach towards efficient deep architectures. In: arXiv preprint arXiv:1607.03250 (2016) [3] LI, Hao ; KADAV, Asim ; DURDANOVIC, Igor ; SAMET, Hanan ; GRAF, Hans P.: Pruning filters for efficient convnets. In: International Conference on Learning Representations (ICLR) (2017) [4] MOLCHANOV, Pavlo ; TYREE, Stephen ; KARRAS, Tero ; AILA, Timo ; KAUTZ, Jan: Pruning convolutional neural networks for resource efficient inference. In: International Conference on Learning Representations (ICLR) (2017)

Clarity: The paper is well written and easy to follow. Some evident typos: - bad-formatting at line 3 of Algorithm 1 - "Postivie" instead of "Positive" in the title of Fig. 1 In general, I would prefer if the comparison with Neural Conductance was inlined in each experimental section, instead of being left in a marginal position.

Relation to Prior Work: Most relevant related works are cited and the differences are discussed, even though some clarification is needed (see above).

Reproducibility: Yes

Additional Feedback: Overall, I find the paper quite interesting, clear and easy to read. The idea of using MAB to find the highest ranking players is nice and I consider this the major contribution of the paper. Shapley values is not a novel idea in the XAI community and it was previously used in the pruning literature for the ranking of hidden neurons, so the novelty is limited in this direction. The experimental section needs better comparison with other methods or at least a fair comparison and discussion of conductance, which should be much faster to compute. ====================================================== After reading the authors' feedback and the other reviews, I increase my score to 7. This is surely an interesting paper. As pointed out by another reviewer, MBA applied to Shapley Values is not entirely new. I still believe the authors should clarify their experiment on adversarial examples as a "black-box attacks" normally refers to a setting where the attacker does not have access to the model parameters but *has* access to the prediction of the model for a given input. In this setting, many "query-based" attacks have been proposed and it does not seem that the experiment in this paper actually runs any of these on the pruned network, thus making the claim potentially misleading.


Review 2

Summary and Contributions: the authors propose using the Shapley value over a game over model elements (neurons/filters or any other type). The value of a subset of neurons can be a lot of things, like model accuracy, model fairness or anything else really. The authors suggest a multi-armed bandit approach to compute the Shapley value (this is useful in applications where one just wants to identify the more important elements rather than knowing their precise Shapley value) and evaluate their approach on a variety of datasets. The results show that removing the neurons identified as important by the approach indeed results in a significant drop in network accuracy. Similarly, removing neurons responsible for bad model behavior results in significant improvement. This is useful since the number of neurons that need to be removed is typically small.

Strengths: This is an interesting application of the Shapley value to ML, which makes sense in the domain. The usage of a multi-armed bandit approach speeds up the computation significantly and this trick might even be useful in other situations where Shapley values are calculated The experimental results, including the potential application to fairness seem promising. The most interesting thing about the paper is the authors experimentally show how successful in different tasks, especially finding out class-specific crucial filters. It’s surprising when the proposed method can find filters that affect only one target class with little changing to the overall accuracy. This will help understand the network better.

Weaknesses: Some of the experimental results need more details and elaboration. Top-k crucial filters: the authors did not yet explain the interaction between these filters as mentioned in line 65. How are these filters related to each other? Are they from different layers or the same layer? Robustness experiment: In line 275, when the authors mention ‘original adversary’, do they mean the PGD attack (i.e. rerunning the attack on the modified network) or just the adversarial images generated beforehand? If the latter, the result is much less surprising. What is strange about the paper is the evaluation method. First, the authors need to estimate Shapley values for individual filters from performance of coalitions. Then for most experiments, they measure the performance of the method by zeroing out the set of top k filters. If we are interested in top k crucial filters, is there any way to find them other than via computing Shapley value? Feature selection methods seem like a good candidate for comparison, not just objective performance. I would like to point out that https://arena.gov.au/assets/2019/06/consort-reward-structures.pdf discusses a similar approach for Shapley value computations using MAB.

Correctness: The claims about the achieved speedup are a bit vague and only evaluated on two datasets. Which seems more like a proof of concept than a thorough investigation.

Clarity: Yes overall, with a few typos and minor issues: see additional comments. There are minor issues with typos and Algorithm 1 is ill-defined. The authors use the terms “neuron” and “filter” interchangeably, which is confusing, and calling the methods “Neuron shapley” feels misleading.

Relation to Prior Work: yes

Reproducibility: Yes

Additional Feedback: line 21-23: sentence fragment, Line 26: neruons -> neurons Line 134 “This, to our knowledge, is one of the few methods that take such interactions into account.” → There are actually several such methods in the Game Theory literature. For example the Banzhaf interaction index (see Hammer and Holtzmann “Approximations of Pseudo-Boolean Functions”) line 137: v: 2^n \to R Line 142 appendix b → Appendix B (similar in other lines) Line 155 literature. [29, 6, 28] → literature [29, 6, 28]. Line 180 our three approximation methods, → counting Monte Carlo as “our” is a bit of a stretch Line 200 This statement is very unclear, given that TMAB already includes the truncation. What does the 11.5 in Line 194 mean is this only from the Bandit approach? Algorithm 1 Line 3: Something went wrong after sigma Algorithm 1 is unclear and/or wrong v is not clearly defined and or initialized line 10-11 make no sense without initialisation Line 274-277 This statement is very vague. Also robustness against a specific attack is practically useless. See “On Evaluating Adversarial Robustness” Carlini et al. 2019, for guidelines on testing adversarial robustness


Review 3

Summary and Contributions: This paper introduces a bandits-based algorithm to quantify the importance of each filter and/or neuron in a network. The authors bridge the idea of the Shapley value from cooperative game theory in order to devise a saliency measure that takes into account the interdependencies/interactions between the neurons, and more generally, components of the network. The authors then use and build upon ideas from bandits literature in their algorithm to tame the intractability of exact Shapley value computations, and to identify the neurons with the highest Shapley values (i.e., most important neurons).

Strengths: - The paper is of high relevance as it attempts to remedy a notable shortcoming of contemporary work on pruning -- capturing the complicated interactions between the weights/neurons/filters of a network when assigning importances to or understanding the role of each neuron/filter - The idea of using a bandits-based algorithm for pruning is novel and very interesting - Nice applications and empirical evaluations to fairness and adversarial attacks, and some compelling results

Weaknesses: - Lack of theoretical guarantees: although the bandits-based algorithm is interesting and a principled approach to this problem, I was not convinced that it was better than MC sampling (with an importance sampling distribution) since it has theoretical bounds on the accuracy of the approximations. The authors state that bandits-based should fare better since most of the Shapley Values will be close to 0 (i.e., most neurons are not that important), but there is no regret bound given for the bandits-based algorithm and Fig. 3 (a) doesn’t prove that the *total* number of samples required by the algorithm is smaller. - Lack of clarity and detail in the method and background: I would have liked to see more of an explicit connection to bandits literature (e.g., the rewards in our case are so and so, the arms correspond to different neurons, the rewards are… and are bounded in [0,1] because...). It is also not clear what the computational cost of the algorithm is and how it compares to MC sampling. Since the algorithm has a while loop, it is not clear whether the terminating condition is always reached or what the (expected) number of iterations is. It is also not clear how the algorithm scales with the input k (the number of neurons to “keep”). Can it be so large as to be used in network pruning so that only the essential filters are kept? Update: I have raised my score to a 6 in light of the reviewer discussion and the authors' rebuttal. I still think that the work can be improved by using existing work in MAB for pure exploration (e.g., see [1]), and I think that trying to describe a properietary algorithm rather than refer to well-established MAB work hurts the exposition and makes it difficult to understand. Given the discussion, I am more convinced that the tools used in this work -- MAB & Shapley values --, although not novel, may be of interest to the pruning community in particular, since the main challenge in pruning is in dealing with parameter interdepencies in a computationally efficient way. I would encourage the authors to emphasize this connection and potential application to pruning in their work. [1] https://tor-lattimore.com/downloads/book/book.pdf , Chapter 33 (pg. 409)

Correctness: My biggest qualm with the claims in this paper is the lack of quantitative results and comparisons to other neuron importance measures, especially those that were developed for use in network pruning (e.g.,[1-3]). The author did note that a future direction is to use their algorithm in a prune-retrain (or change-retrain) pipeline, which is fine, but even a simple prune-only experiment that demonstrates the value of the bandits-based approach with Shapley values in retaining only the important filters of a network would make the paper much more impactful

Clarity: The paper is well-written for the most part, but I would have appreciated more details and clarity in the description of the method and the pseudocode.

Relation to Prior Work: For the most part, there is a nice exposition of related work and the authors justify each component of their algorithm by relating it to prior work. However, some claims in relation to previous work, such as the superiority of bandits-based vs. MC sampling and the empirical or theoretical accuracy of the Shapley Value in contrast to previously introduced neuron importance metrics could have used more justification (either theoretically or experimentally)

Reproducibility: Yes

Additional Feedback: