Alignment with human representations supports robust few-shot learning

Should we care whether AI systems have representations of the world that are similar to those of humans? We provide an information-theoretic analysis that suggests that there should be a U-shaped relationship between the degree of representational alignment with humans and performance on few-shot learning tasks. We confirm this prediction empirically, finding such a relationship in an analysis of the performance of 491 computer vision models. We also show that highly-aligned models are more robust to both natural adversarial attacks and domain shifts. Our results suggest that human-alignment is often a sufficient, but not necessary, condition for models to make effective use of limited data, be robust, and generalize well.

Paper

References (43)

Scroll for more · 31 remaining

Similar papers

Peer review

Reviewer 225s6/10 · confidence 3/52023-07-03

Summary

This paper explores the relationship between the alignment of deep learning based representations with humans and the downstream performance of the NN models. Using triplets as a non-metric measure of similarity, the authors use information theory to show that alignment and downstream performance would have a U-shaped relationship, where low and high alignment would give a good downstream performance.

Strengths

* The authors use existing ideas of triplets from cognitive science and combine information theory to show why human alignment may not always lead to better performance, especially for few-shot downstream learning. More importantly, the authors give an intuition and theoretical backing for why some models can have low alignment with humans and still perform well in downstream learning. * The authors highlight the significance of ensuring human alignment of neural network representations by studying the role of alignment on two critical characteristics: adversarial robustness and domain shift robustness. They further show how human alignment plays a role in downstream learning, mainly when limited data is available. * Using a significantly large number of models and large-scale existing human-annotated data, the authors empirically validated the theoretical claims made in the paper, showing the benefit of very high alignment with robustness while also revealing the U-shaped relationship between few-shot learning and alignment. The authors performed critical statistical testing to highlight the results. * The interaction of architecture model choices and alignment is insightful and could be helpful for future work.

Weaknesses

* The contribution seems limited, the main contribution seems to be the showing of the U-shaped relationship. The primary notion of triplets is derived from prior work. * The importance of triplets seems limited. Pearson and Spearman also clearly provide similar takeaways. Why are triplets needed? Not clear from the evaluation section. * Evaluation is also limited in some aspects, although the paper explores more than 400 models. Most transfer learning results are shown for CIFAR after ImageNet pre-training. The result on MNIST and FMNIST show quite weak results, meaning that strong domain shift during transfer learning makes the metric less impactful. But transfer learning is usually most important in these settings (e.g., medical domain, face classification, etc.). * There is a lack of clarity in the writing of the paper. Section 3 is difficult to follow, and it is not always clear how the different definitions flow, and how the pipeline works. E.g., not clear what is the significance of Def. 3.4, Discussion section mentions "quadruplets", but they are never mentioned in the paper. Not clear at all. * Adversarial robustness is measured for real-world adversarial images. More common adversarial examples used in literature, e.g., $\ell_p$ ball perturbations are not explored. * Evaluation section compares the triplets with pairwise measures like Pearson and Spearman correlations. But it is not clear from Section 3 as to why it is interesting to compare triplets to pairwise. Do triplets give benefits over Pearson/Spearman? Are there any tradeoffs? * Results shown in the tables are not well discussed * Transfer learning results only analyzed when we have "head" networks. Not clear if the method works for full network fine-tuning or not, something that is often used in practice. * No justification given on the choice of the heads, e.g., dropout rate, etc. Were they hyper parameter-tuned? * Not clear what and how the ImageNet accuracy correction is performed. It seems important, so more intuition should be provided.

Questions

* Why does correlation with test accuracy reduce with more training examples (Fig. 3 and 4 right)? * How are the classifier heads parameters chosen? * What is meant by supervised learning in Section 3? Does it mean standard classification task? It is not clear. * What are quadruplets mentioned in Discussion? * Large datasets like ImageNet and CIFAR have a lot of "noisy" and misclassified examples on which the deep models are trained on. Do the different alignment metrics reveal any interesting facts on the model's alignment or non-alignment on these "outlier" examples?

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

* Most transfer learning results are shown for transfer from ImageNet to CIFAR, which is limiting. * Exact contribution of triplets is not clear from the paper, as existing correlation metrics can already show the U-shaped relation and performance relations.

Reviewer wjPt6/10 · confidence 4/52023-07-05

Summary

The work provides theoretical and empirical analysis for the relationship between human alignment, few-shot performance, and robustness to distribution shift. The theoretical analysis predicts models that have high and low alignment with human annotations will have better performance on few-shot and distribution shift datasets. The experiments verify the predicted relationship between representational alignment and downstream performance.

Strengths

- Alignment between humans and ML models has been studied considerably from the perspective of interpretability and safety. Understanding how and whether alignment can improve sample-efficiency and robustness is a less explored, but important research question which this work makes progress towards. - The paper theoretically and empirically shows that alignment is not monotonic with respect to robustness/sample efficiency which may explain why previous works have found conflicting results regarding the benefits of alignments. I think this finding would be useful to the community. Also the information-theoretic framework proposed by the authors for analyzing representational alignment would likely be useful to future work. - The experiments are thorough and adequately support the claims. - The paper is well organized and the tables and plots clearly illustrate the main empirical findings.

Weaknesses

Other works [1][2] have shown that the validation accuracy on an in-distribution (ID) dataset is highly correlated with accuracy on out-of-distribution (OOD) and downstream datasets. For reference, the R-squared value for pretrained models evaluated on ImageNet and ImageNetV2 is .996 and other distribution shifts have similar correlation. Given that in-distribution validation accuracy is almost entirely predictive of out-of-distribution accuracy, I think the current motivation, which is to show that alignment is predictive of downstream performance, is weak. Representational alignment as a predictive metric for robustness and downstream transfer performance is less accurate than standard ID validation accuracy. Also obtaining validation accuracy is easier since it only requires one-hot labeling, as opposed to similarity matrices. If the main motivation of the paper is to show that improving alignment will confer benefits such as robustness and few-shot performance, then it should show that alignment measures something more than in-distribution validation accuracy. Otherwise, robustness and few-shot performance could simply be increased by improving the model on the validation set. Overall I think the paper is interesting and original. My main concern is for the motivation, but would be happy to change my rating if this concern is adequately addressed. [1] Accuracy on the Line: On the Strong Correlation Between Out-of-Distribution and In-Distribution Generalization [2] Do ImageNet Classifiers Generalize to ImageNet?

Questions

- It would be helpful to see the distribution of alignment metrics and some analysis on what factors other than validation accuracy affect the alignment metric. Given that validation accuracy is almost entirely predictive of robustness and downstream performance (greater than .9 R-squared), it seems that validation accuracy and alignment would be highly correlated. - In the abstract it’s stated that highly-aligned models are more robust to adversarial attacks. This claim seems overly strong as the experiments are conducted with natural adversarial images, and not with white or black box attacks. - Given theorem 3.9, would the relationship between alignment and few-shot performance be v-shaped and not u-shaped?

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately address the limitations of the method.

Reviewer kVdE7/10 · confidence 4/52023-07-07

Summary

In this paper, the authors present a theory for a u-shaped effect in terms of model representation alignment, where models that are more and less aligned with human representations should have better performance over a number of tasks. They then perform a large number of experiments, comparing 491 models with a representation approximated based on 425,000 human judgements. They find that the quantitative results largely hold with their theory. Edit: Based on the author's rebuttal and further discussion I have raised my initial review to an accept.

Strengths

This paper has a large number of strengths. It is well-written and well-presented. The theory is novel and valuable, providing insights for future researchers and developers. The experiments are exhaustive and well-motivated. The depth of discussion around justifying the choices in the experiments is excellent. The experiments also largely follow the theory, which suggests useful takeaways to a broad audience.

Weaknesses

This is a strong paper, but there's a couple of weaknesses which dampen my enthusiasm somewhat. The largest of these weaknesses is that while the authors repeatedly refer to a U-shaped effect, the results actually suggest a J-shaped effect. The low alignment models consistently demonstrate a significantly worse performance than the high-alignment models Despite this, there's no discussion of this point throughout the paper. The reasoning seems intuitively straightforward, despite note following from the theory. While it is true that a model could "realize" to interpret signals in an "opposite" manner (to use the paper's term) with low alignment, this realization is still extra effort. In a model trained on backprop, it is not always possible to "flip" the use of features in a large, multidimensional space. This or similar discussion accounting for this issue should be brought up in the paper. Second, while a large number of experiments are run, the choice of only a top 5, bottom 5, and mean 5 is somewhat odd. This represents less than 3% of the models used in the experiments. This is understandable in some cases, like presenting individual values in tables. But in Figures like Figure 3 a more continual effect could have been demonstrated with a typical standard deviation off the lines. If the authors chose to only use 15 models, some experiment helping to demonstrate that the results weren't a fluke would be beneficial. For example, ranking the 491 models in term of alignment and visualizing some performance metric across all 491. As it is, readers need to look through a dense CSV file for this information. There's notably no readMe or similar to explain the labels for the CSV file making it difficult to understand what's going on here without digging through code. Some clarity on this would be appreciated and would benefit the paper.

Questions

1. How do the authors account for the effect being more J-shaped than U-shaped as predicted? 2. Why did the authors choose to only look at high-5/meanest-5/lowest-5 models?

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 authors identify the issues around the choice of alignment, otherwise there are no major potential negative societal issues.

Reviewer CS1b7/10 · confidence 3/52023-07-27

Summary

- The authors predict that there should be an U-shaped relationship between the degree of representational alignment and performance on few-shot tasks based on information theoretic analysis, and empirically show that is the case using hundreds of vision models. They further show that the degree of alignment predicts robustness to ImageNet-A, domain-shift (e.g. ImageNet-R, -S.)

Strengths

- The U-shaped results are novel and interesting phenomenon. With both information theoretical analysis and detailed empirical analysis, the results seem credible, and provide new perspective on the previously conflicting findings as to whether improving alignment increases performance. - They study three different alignment metrics. - The degree of alignment not only explains the few-shot performance, but also predicts robustness to distributional shift (e.g. ImageNet-A, -R, -S) which intuitively make sense.

Weaknesses

- Minor: - I believe ImageNet-A is just a subset of ImageNet-test set that happens to be difficult, but the term "adversarial robustness" often refers to images with actual adversarial noise, so I think it should be clarified.

Questions

- Figure 4 Left: Curious to see if there is any characteristics of models that are in the top left of the graph (the ones with test accuracy around 60% and alignment being less than 20), as these points seem to be rather outliers.

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

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

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

- The analysis seems to be limited to ImageNet-based models. Recently there are more capable vision-language models such as BLIP-2, which combines a vision model with Large Language Model. While the current paper conducts detailed analysis, it would be a more interesting work if they also study these recent vision-language models.

Area Chair XxdF2023-08-10

Author-Reviewer Discussion phase (Aug 10-16)

Today begins the Author-Reviewer Discussion phase, which lasts 1 week (**Aug 10-16**). I ask the reviewers to please **carefully read all other reviews and the author responses and (if appropriate) respond to author responses promptly.** If you've read the author response, please take the time to leave a comment, even if you have nothing to add. I also encourage both authors and reviewers to monitor OpenReview for further comments in order to enable as much back-and-forth as possible during this short period. As the only reviewer recommending rejection, I especially encourage Reviewer 225s to engage during this period.

Reviewer 225s2023-08-10

I thank the authors for the detailed rebuttal and comments addressing all the reviewers' concerns. The authors have clarified the doubts and addressed my concerns adequately. Hence I improve my prior rating and suggest acceptance of the paper. The work is interesting, potentially showing the reason behind some prior conflicting results, and opens the avenue to further research in this direction. I especially find the results on the ImageNet-ReaL dataset fascinating and show a need to revisit ImageNet to study the representations and their alignment.

Reviewer kVdE2023-08-11

Re: Rebuttal by Authors

Thanks to the authors for their comprehensive response! C1: This is an excellent explanation and definitely fits better than my rough intuition. I think the inclusion of it greatly improves the paper. C2: I apologize for misunderstanding those later figures and tables! I appreciate the clarification. Based on the improved discussion and my improved understanding of the work I have upgraded my score to an Accept.

Reviewer wjPt2023-08-14

Response to Author Rebuttal

I thank the authors for their response. My concerns have been addressed, in particular the motivation for the work and the correlation between validation accuracy and robustness. I have raised my score accordingly.

Reviewer CS1b2023-08-18

Comment

Thank you for pointing out my misunderstanding / clarifying the figure. The authors have addressed my questions and suggestions. I think this paper will be a great contribution to the community. I keep my score (7: Accept).

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC