Evaluating alignment between humans and neural network representations in image-based learning tasks

Summary

This article assesses how AI embedding can predict human behavior on 2 different tasks. In those tasks, humans have to learn continuous relationships and categories of natural images. Overall the authors tested the predictability of 77 retrained neural networks, including supervised, semi-supervised, and multimodal models. The authors found that a larger training size, a higher number of parameters, and lower intrinsic dimensionality improve the human alignment of the tested model. In addition, the authors claim that multimodal networks are better aligned with humans. In the last part, the authors test 3 different alignment methods and show that only the gLocal method shows a better predictability of human performance on the considered task.

Strengths

This article introduces a novel behavioral benchmark to compare humans and machines. This benchmark (provided that the data are public) should be a useful asset in the human-machine comparison toolbox. The number of tested models is both large and also diverse.

Weaknesses

I found 3 weaknesses: * 1 - Given the data presented in the article, I have the feeling that claiming that ‘multimodality better predicts human performance’ is a bit overstated. It seems to have plenty of confound factors (besides the training side) that haven’t been explored and that might prove that this claim is not entirely true * 2 - Some of the parts in the article are very unclear with several details missing. I have the feeling I did not have clear enough to judge the methods used by the authors (which is problematic). Please refer to the question section for more details. * 3 - The last analysis (on the alignment method) is pretty weak compared to the rest of the article (not enough models and not enough model diversity, no statistical analysis…). Note that I am willing to increase my rating if the questions are properly addressed.

Questions

* Q1: The section 3 is not clear enough. For example (line 95): « Using an intercept and the trial number ». What does it mean? To my (maybe not sufficient) knowledge, the intercept is the constant term in a linear fit. If this is the case, what kind of data are you using to regress to find the fit? And most importantly, how does this intercept relate to human performance? * Q2: Still in section 3, the results of the statistical tests are poorly explained. For example (line 98): what does (Beta, z) relate to? Is that the intercept and the number of trials? Is that something else? This should be explicitly stated and should be put under the carpet! * Q3: In section 4, there is not enough detail on how you train the linear probe to predict the performance of the category learning and the reward learning task. Especially in the case of the ‘linear regression mode with spherical prior » (line 121). What does it mean, what is the loss you exactly optimize to find the regression parameters? Why Gaussian prior? This should be motivated and better explained... * Q4: Even more dark for me is line 171. T is the generative task representation. What does it mean, and how this is obtained? What is the generative task? What’s the experimental protocol for this task (if there is one)… * Q5: I am lost in Figure 5. If I am not wrong clip is represented by the yellow color (multimodal, as confirmed in Fig 3). But in the caption of the same figure you say: « lower intrinsic dimensionality only increases alignment for CLIP models ». But the supervised model (red bar) showcases a similar trend. Is that a typo? Do I miss something here? * Q6: In Figures 5, 6, there is no error bar and statistical test to assess if the shown results are significant. For example in Fig6, the differences between CLIP, SimCLR, and CLIP+SimCLR seem to be rather small and I cannot evaluate if those differences are significant. * Q7: In Figure 3, would it be possible to include inter-human reliability (i.e. how much different participants are aligned with each other) * Q8: Is it not clear how to control the number of training data in CLIP and SimCLR? In CLIP, this is pairs of data (language/image) whereas SimCLR is only on images (and it’s an augmented version, but it might not count as it is not external information per say). * Q9: The authors suggest that the training size is a confound (and I agree), but the model’s number of parameters might also be a confound (that might be related to the training size). This is a problem because most of the multimodal models have generally a high number of parameters (see Fig B). How do I know if this is multimodality is the number of parameters that increase the human alignment? Would it be possible to run a similar experiment as in Fig 6, but testing the number of parameters? * Q10: I have the feeling there is no enough data (and a too small variety of tested model) to back the claim : « We found that only one alignement method improved predictive accuracy » (line 15). For example, for the harmonization technique only supervised models have been tested, and several other models have been tested with gLocal. How do you know that « harmonising » other type of model do not improve the predictability on your task. And I am even sure that the tested architectures are the same for the supervised harmonized and supervised gLocal models ? How do you want me to be convinced if on top of the alignement methods, other parameters can vary (as architecture…). This analysis clearly lack rigor...

Rating

5

Confidence

3

Soundness

2

Presentation

2

Contribution

3

Limitations

The authors properly discuss the limitation of their work

Summary

This paper explores the alignment between human cognitive processes and neural network representations in image-based learning tasks. The authors evaluated 77 pretrained neural network models to see how their representations aligned with human learning patterns in two tasks: category learning and reward learning. They found that models with contrastive training, especially CLIP models, effectively predicted human generalization. Factors such as training dataset size, model size, and representation properties were identified as influential for alignment with human cognition.

Strengths

The paper makes substantial contributions by: 1. Introducing two novel tasks to evaluate human-model alignment. 2. Demonstrating that larger contrastive training datasets and multi-modal data improve predictions of human behavior. 3. Conducting extensive and engaging human experiments. Additionally, I found that: - Figure 4 is particularly interesting and stimulates a lot of valuable discussion. - It would have been very insightful to include a comparison with generative models (e.g., EBM, VAE) and relate these findings to the brain's generative hypothesis.

Weaknesses

Despite its strengths, the paper has several weaknesses. These are categorized into major problems (**M**) and minor problems (_m_). **M1**: There is a potential bias in using the THINGS database for both training and evaluation, especially with the gLocal models, which might lead to circular reasoning. The space trained with gLocal is implicitly designed to perform well on these specific tasks, which could inflate the perceived performance of the models. Addressing this concern by diversifying the evaluation datasets or providing a more detailed justification for using THINGS would strengthen the paper. You could use a measure from DreamSim and Harmonization as control. _m1_: The use of $\ell_2$ regularization in the regression model may penalize some tokens disproportionately due to their high activation values. Standardizing all activations before performing regression could mitigate this issue and provide a fairer comparison. _m2_: Concerning the use of the TwoNN method for estimating intrinsic dimensionality, could you verify that standardization instead of linear scaling yield the same results, as we know that some high norm token (outlier token) exist and could totally distort your distances. _m3_: Figure 7 lacks clarity on which models are used in the harmonization method? You have 7 Harmonization model in Fig. 3 and only 3 in Fig. 7. Or at least specify how they were chosen. _m4_: Could you also add a metric to complement McFadden (e.g., avg accuracy) ?

Questions

See Strengths & Weaknesses

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, the limitations identified by the authors are accurate and well-documented. Regarding the weakness I mentioned, I reserve the right to increase the score if the authors adequately address my major concerns.

Summary

The authors propose a novel representational alignment metric tied to sequential human behavior and leverage it to analyze factors in NN design and training that contribute to increased alignment with humans.

Strengths

- Great analysis of factors contributing to alignment of NNs with humans - Well-written manuscript, easy to follow - Great descriptions of human and NN experiments; important for reproducibility - Important and timely contribution to a growing field (representational alignment) - I agree with the authors that its important to tie notions of alignment to behavior (a metric is only useful if it actually predicts something downstream)

Weaknesses

- The authors claim to make 2 contributions: a novel alignment metric and an analysis of factors contributing to human alignment of various NN that leverages that alignment metric. The authors claim that their alignment metric is better than existing alignment metrics as it requires "generalisation and information integration across an extended horizon". However, there isn't a comprehensive comparison to previous alignment metrics to see (a) how correlated the new metric is to existing metrics, and (b) whether human behavior on the tasks can be predicted from previous metrics. For (b) for example, the authors could take human pairwise similarity judgments and treat them as features from which to learn the task and then correlate that with human responses (i.e., treating the sim judgments as a kernel). This kind of analysis would go a long way in validating the proposed metric and comparing/contrasting it against existing alignment metrics. - Perhaps I missed it, but I didn't see an analysis of inter-rater reliability/noise roof. This would be important to put the McFadden R^2 numbers into perspective and understand how good alignment *could* be. - Happy to raise my score if these points and questions below are resolved

Questions

- why were participants with <50% accuracy excluded? - why might the alignment numbers be so low? McFadden's R^2 is typically considered to be excellent fit in 0.2-0.4 range, yet all the results cap out at about 0.15

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

- Limitations are adequately discussed (assuming weaknesses and questions raised above are addressed)

Summary

The authors evaluate alignment between humans and 77 pre-trained neural network vision models using learning tasks. Previous work has focused on comparing alignment between humans and models with similarity judgments alone; instead, here the authors asked participants to perform simple category learning and reinforcement learning tasks. These responses were compared with linear models fit on top of the pre-trained networks. The analysis is very thoughtful and thorough, and the article goes into detail regarding the factors that influence which models are best aligned: task accuracy, model size, training, etc. I see this work as providing a strong contribution to understanding human vs. model alignment in vision tasks. I would like to see it published at NeurIPS, pending the concern I mentioned below.

Strengths

This article has many strengths - Kudos for comparing 77 different models - Thoughtful and rigorous model comparison - Detailed analysis of which factors lead models to fit the human data better - Excellent visuals for understanding the results

Weaknesses

My main concern is about the method for fitting participant responses (see question below). The categorization task of delivering images to two dinosaurs, Julty and Folty, is a bit silly and may have had participants overthinking the task. You may have received cleaner data with a simpler framing: learning Category A vs. Category B. But I don't see this as a major issue.

Questions

I don't fully understand the methodological choices in getting the pre-trained models to predict the human responses. In the categorization task, my understanding is that a regularized logistic regression was trained to predict the gold category labels. Then, the predicted category probability was used in a logistic mixed model. With this approach, you would seemingly lose a lot of information (potentially relevant features) in the model embeddings beyond its category prediction, before mapping this prediction to human responses. Did you consider fitting a mapping between the model features and participant responses more directly? If not, why not? I would be willing to raise my score depending on the response to this question. EDIT: Thank you for trying this and for the explanation. I see the current approach as better justified now, and raised my score accordingly. Also, the description of the mixed model in the appendix is not satisfying, as the variables aren't defined clearly. Please explain in the rebuttal. Typos "Exemplary learning curves" (pg 4)

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

This section is fine

response

I still think the claim of this article is overstated, and even if results tend to show an interesting trend, there is still many confound that has not been explored. I would have appreciate to downplay the overclaim (and to discuss more the other possible confound). I maintain my rating.

Discussion Period Summary

Dear all, We thank all reviewers for their engagement throughout the review process. **All reviewers recommend acceptance, with an average of 6.** After our rebuttal, three reviewers [pzQG, MCn8, NCLd] increased their scores. We appreciate the reviewers' constructive input and their recognition of these improvements in their updated evaluations. Thank you all again.