Towards a fuller understanding of neurons with Clustered Compositional Explanations

Compositional Explanations is a method for identifying logical formulas of concepts that approximate the neurons' behavior. However, these explanations are linked to the small spectrum of neuron activations (i.e., the highest ones) used to check the alignment, thus lacking completeness. In this paper, we propose a generalization, called Clustered Compositional Explanations, that combines Compositional Explanations with clustering and a novel search heuristic to approximate a broader spectrum of the neurons' behavior. We define and address the problems connected to the application of these methods to multiple ranges of activations, analyze the insights retrievable by using our algorithm, and propose desiderata qualities that can be used to study the explanations returned by different algorithms.

Paper

References (46)

Scroll for more · 34 remaining

Similar papers

Peer review

Reviewer 4jti2/10 · confidence 5/52023-06-13

Summary

This paper is a niche extension of seminal work on network dissection. The authors present a generalization, called Clustered Compositional Explanations, that combines Compositional Explanations with clustering and a novel search heuristic to approximate a broader spectrum of the neuron behavior.

Strengths

Easy to read

Weaknesses

cf. Limitations

Questions

N/A

Rating

2: Strong Reject: For instance, a paper with major technical flaws, and/or poor evaluation, limited impact, poor reproducibility and mostly unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

1 poor

Limitations

This is mainly a search-based approach on a set of clusters which group neurons. This is very incremental work. Research angle of the work is limited to the application of search on clusters grouping neurons No novelty

Reviewer 2vbF6/10 · confidence 4/52023-07-04

Summary

The authors propose a novel XAI method called Clustered Compositional Explanations (CCE), that aims to descibe the function that a group of neurons in a neural network perform. The method is built on top of CoEx (Mu and Andreas 2020) and NetDissect (Bau et al 2017) with the novelty being its generalization to multiple activation ranges instead of high activation thresholds only. The objective for CCE aims to maximize the IoU score between the activations of an input and a neuron in the range of min and maximum activation thretholds across all clusters, over all set of logical connections between the annotated sets of concepts in the dataset. The authors solve the optimization problem by presenting a heuristic algorithm MMESH. Results from the authors indicate that the explanations obtained from proposed method does better than NetDissect and CoEx on various metrics that include accuracy and coverage among others.

Strengths

The connection between Clustered Compositional Explanations and existing neuronal explainability methods CoEx and NetDissect is explained very well. The proposed method generalizes both these methods, and the reviewer agrees with the authors that explaining groups of neurons over larger activation ranges could lead to broader explanations. This is confirmed in the results, as the proposed CCE method outperforms CoEx and NetDissect on various metrics, but especially so on coverage related metrics.

Weaknesses

1. My biggest concern is with the MMESH algorithm, and while it is admissible as shown in the appendix, the algorithm is tailored to the CCE method and not easy to follow. Unless the authors plan to release their code (not included in submission), I believe it will greatly limit the utility of their approach since their is a higher barrier of implementation for practitioners versus CoEx. 2. I would like to see results across a broader range of models and datasets in order to ensure that the proposed method does not overfit to explaining the neurons of a particular model, dataset, and training procedure. I think this is reasonable since MMESH takes less than 2 minutes per unit (L219). All results have been reported on CNNs (AlexNet, ResNet, DenseNet) which I presume are trained with supervised learning.

Questions

1. Could the authors please explain why having more clusters than one is necessary, since in Table 2 the first cluster has perfect dataset and sample coverage and outperforms the averaged across clusters on all desiderata, similarly for Table1, 3 in the appendix.

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

4 excellent

Contribution

2 fair

Limitations

The authors have both mentioned potential limitations and highlighted avenues for future work. See also Weaknesses.

Reviewer CMC66/10 · confidence 5/52023-07-04

Summary

This paper extends the ideas of Mu (2020) to examine a more powerful class of compositional explanations of neurons, by adding the goal of explaining other ranges of neuron activations, unlike previous work that had restricted analysis to the top ranges only. Like the previous work by Mu, the paper searches for compositions of human-understandable concepts that explain a neuron’s behavior, but the proposed method begins by first subdividing neuron activations into ranges. Adding the ability to explain lower ranges results in much better coverage: it allows the authors to create explanations of a larger portion of neurons’ behavior, and as a result this paper achieves much higher coverage metrics, explaining more neurons with high matching scores, compared to previous methods.

Strengths

The main strength of this paper is the way that it brings a systematic interpretability analysis to neurons and ranges of neuron activations that have not previously been systematically analyzed. The paper finds that, at lower activation ranges, consistent interpretations can be found for most neurons. On the other hand, the paper finds that most of these interpretations are generic concepts like sets of colors, and the paper hypothesizes that these observed labels are “default labels” that describe inputs when neurons are in a random state. The paper applies their explanations to neurons in a random untrained network to validate this hypothesis; that is a good baseline-setting experiment. The paper also finds that as neuron activations rise to higher levels, the explanations of their behavior becomes progressively more specific. Unsurprisingly, when analyzing the entire activation range of neurons, the paper finds that more individual neurons activate on several unrelated concepts over their range, observing a higher level of such polysemanticity than previous works, but interestingly, the paper finds a portion (15%) of neurons that have a consistent semantics over their range.

Weaknesses

On weaknesses: beyond the clever uninitialized-network experiment, the paper does not conduct experiments that would triangulate the proposal that the intermediate-activation states of neurons might be meaningful. For example, it is natural to ask: if there are ranges of concepts that seem to match middle-range activations, is the network actually "looking" for those midrange concepts? For example, will forcing those neurons to those middle ranges cause a network to change its predictions towards classes that correspond to those explained mid-range concepts? The paper makes a good contribution by increasing the observational breadth of neuron interpretations to include middle-range activations, but it does not directly defend the idea that such midrange activations will be useful to understand.

Questions

The main question is the one posed above. How can we know that the concepts matched in middle ranges of activations are meaningful to map out? Specifically: the paper observes that that the neurons sometimes match different concepts in the middle ranges than they do at extreme ranges. Do the mid-range matched concept correspond to meaningful decisions by the network? For example, would forcing neurons to those middle ranges cause the network to behave as if those middle-range concepts are detected? Does it cause the network to behave as if different concepts were detected compared to those same neurons set to the highest ranges or lowest ranges?

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors note several of the potential limitations of their methods and address some of them in the appendix. No immediate negative societal impacts are anticipated from this line of work.

Reviewer rkdS6/10 · confidence 3/52023-07-22

Summary

This paper focuses on a problem with Network Dissection and Compositional Explanation methods: these two methods explain the concept encoded by a neuron (or, more precisely, a convolutional filter) by only considering highly activated regions in the feature map. To address the problem, this paper proposes to divide the activation values into different ranges (clusters) and generate an explanation for each range of activation value. This paper also proposes a heuristic to accelerate the search for the optimal logical combinations of concepts in the Compositional Explanation method.

Strengths

1. This paper focuses on an important issue, i.e., previous concept-based explanation methods, such as Network Dissection or Compositional Explanation, neglect feature map regions with low activation values. 2. The authors theoretically prove that the proposed heuristic is admissible, which guarantees the optimality of the solution found using the heuristic.

Weaknesses

1. The contributions are incremental and miscellaneous. From my view, the contribution of this paper is mainly two-fold. The first contribution is that the paper considers different ranges of activation values to address the problem with previous explanations (Network Dissection and Compositional Explanation) that they only analyze the feature map regions with extremely high activation values. However, the proposed method is only a simple extension of previous methods, and the motivation of using clustering instead of manually dividing the activation values into different ranges is not well explained. The second contribution is to propose a heuristic searching method to accelerate the Compositional Explanation method. However, the second contribution is not inherently correlated with the first one, and I feel the paper is separated into two uncorrelated parts. 2. Many of the notations are inconsistent or confusing. For example: (a) In Line 106, the intersection size is defined as $IMS(x,L)$. But in Eq. (5) to Eq. (7), it is denoted as $IMS_{[\tau\_1, \tau\_2]}(x,L)$. The meaning of the subscripts needs to be clarified. (b) In Eq. (1), $\mathcal{L}^n$ should be $L$, in order to be consistent with Eq. (2). (c) In Eq. (5)-(7), there is a notation $L_L$. I’m confused about what the two L mean respectively. If they refer to different meanings, please use different letters. (d) In Eq. (5)-(7), there is no explanation for the subscript $R$ in the notation $L_R$. (e) In many equations, such as Eq. (2), Eq. (11), Eq. (13), Eq. (14), an equal sign is missing. 3. The metric Activation Coverage seems redundant. Given the metric Intersection Over Union and the metric Detection Accuracy, we can actually derive the metric Activation Coverage by using the inclusion-exclusion principle $|A\cup B|=|A|+|B|-|A\cap B|$. Therefore, I doubt this metric is redundant.

Questions

1. In Table 2, when computing the desiderata qualities (or metric scores) of the proposed method, the authors simply average the scores from the 5 clusters. However, since different clusters correspond to different ranges of activation values, does it make more sense to assign different weights to the scores of these clusters? For example, it is natural to assign higher weights to clusters with higher activation values. Nevertheless, the authors are also encouraged to explain why doing simple averaging is a reasonable choice. 2. As noted by the authors in Line 259 to Line 276, the concept labels obtained from Cluster 0 and Cluster 1 usually represent the “default labels” to which the algorithm converges when the activations are random. This implies that the labels obtained from Cluster 0 and Cluster 1 actually cannot explain what the model has learned from the data. In this way, the metric scores of Cluster 0 and Cluster 1 are meaningless and should not be used to compute the final score of the proposed method. 3. Why do the authors set the number of clusters to 5? If there are more or fewer clusters, will the analysis and conclusion in Sections 4.3 and 4.4 change? I suggest the authors conduct ablation studies regarding the number of clusters and clearly explain how it will affect the empirical results. 4. In Table 3, why do the results with ReLU greatly differ from results without ReLU? Could the authors give some explanations or discussions on this issue?

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Yes, the authors have discussed the limitations and broader impact of their work.

Reviewer BxbC6/10 · confidence 2/52023-07-24

Summary

The paper represents a generalization of compositional explanation called clustered compositional explanations which combines compositional explanations with clustering and a search heuristic to approximate a broader spectrum of the neuron behavior, by proposing the Min-Max Extension per Sample Heuristic (MMESH). This paper gives an analysis of the phenomena connected to the neuron's activations like the unspecialization of the lowest activations in ReLU networks and the progressive specialization.

Strengths

1. This paper well delivers its contribution on the generalization of CoEx based on a heuristic and a wider spectrum of activations by clustering them. 3. The experiments are thoroughly conducted with a detailed analysis of the proposed MMESH. The paper also addresses limitations with future direction of the research.

Weaknesses

1. It would be nicer to clearly show the difference between Mu and Andreas [24].

Questions

Please refer to the weakness section.

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Please refer to the weakness section.

Reviewer 2vbF2023-08-14

Response to rebuttal

Thank you for taking the time to write a thorough rebuttal, as well as for adding the experiments related to testing explainability on ImageNet. I think it makes the work more comprehensive, and the authors' release of the CoEx code should make it easier for others to use their method as well. I have therefore revised my rating from a Borderline Accept to a Weak Accept.

Reviewer rkdS2023-08-16

Since my major concerns are addressed by the authors' rebuttal, I would like to raise my rating to 6.

Reviewer CMC62023-08-20

Thank you for the added measurements of prediction changes across ranges.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC