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

(...) claiming that 'multimodality better predicts human performance' is a bit overstated.

We agree that many factors contribute to alignment, as detailed in Figure 4. However, the comparison of SLIP, SimCLR, and CLIP models in Figure 6 suggests image language training is important, as architecture and dataset are matched and downstream performance is comparable [1]. We acknowledge there may be unexplored factors and welcome suggestions for additional comparisons.

Q1: The section 3 is not clear enough.

For category learning, we used a mixed-effects logistic model to predict correct responses per trial. Main effects included intercept and trial number, where the intercept models above chance performance and the trial number models learning. For reward learning, we predicted which image was chosen using a similar model with trial number, reward difference, and their interaction as main effects. Here, the interaction captures the learning effect. The same variables were also used as random effects for individual differences.

Q2: (...) For example (line 98): what does (Beta, z) relate to? Is that the intercept and the number of trials?

The first set of numbers are for the intercept and the second for the trial number. $\hat{\beta}$ is the estimated regression coefficient. z is the test statistic comparing predictor coefficients against their standard errors to test the null hypothesis that the predictor’s coefficient is 0.

Q3: In section 4, there is not enough detail on how you train the linear probe

We added the following details to the Appendix:

"For the category learning task, we used an $\ell_2$ regularised logistic regression model. We relied on scikit-learn’s LogisticRegression class which optimizes the following objective:

$\mathbf{w}^* = \underset{\mathbf{w}}{argmax} \sum_{i=1}^N -c_i \log(p(c_i | \mathbf{x}_i, \mathbf{w})) - (1-c_i) \log(1-\log(p(c_i | \mathbf{x}_i, \mathbf{w})) + \dfrac{1}{2} ||\mathbf{w} ||^2_2$

For the reward learning task, we used a Bayesian linear regression model to infer a posterior distribution over regression weights. We relied on scikit-learn’s BayesianRidge class which infers a posterior distribution assuming spherical Gaussian priors (i.e., $p(\mathbf{w}) = \mathcal{N}(0, \lambda^{-1}\mathbf{I})$) and Gaussian likelihood (i.e., $p(y_i | \mathbf{x}_i, \mathbf{w}) = \mathcal{N}(\mathbf{w}^{\top}\mathbf{x}_i, \beta^{-1})$). Here, posterior distribution can be computed in closed form:

$p(\mathbf{w} | \mathbf{X}, \mathbf{y}) = \mathcal{N}(\mathbf{m}_N, \mathbf{S}_N)$ $\mathbf{m}_N = \beta \mathbf{S}_N \mathbf{X}^{\top} \mathbf{y}$ $\mathbf{S}^{-1}_N = \lambda \mathbf{I} + \beta \mathbf{X}^{\top}\mathbf{X}$

where $\mathbf{X}$ and $\mathbf{y}$ denote the stacked inputs and targets respectively.

We run both models from scratch on each trial using all previously observed input-target pairs. The choice of these models was motivated by previous investigations in similar but low-dimensional settings [2,3,4].“

Q4: What is the generative task?

T is the ground truth of the task, obtained by [5]. They trained a low-dimensional sparse embedding model on the odd-one-out similarity judgements on THINGS. For both tasks, participants were assigned to conditions defined over one of the top three embedding dimensions (Figure 1C). Please see lines 72-91 for a detailed account.

Q6: In Figures 5, 6, there is no error bar and statistical test to assess if the shown results are significant.

Please see Rebuttal Figures 5 & 6. In Figure 5, we only report the intrinsic dimensionality analysis because the rest are not significant, which we earlier pointed in Appendix Table 3. We hope this also addresses Q5. For Figure 6, we found CLIP $R^2$ values were higher than SimCLR $R^2$ values ($\hat{\beta} = .014, t=1.98, p=.048$), and SimCLR + CLIP $R^2$ values were higher than CLIP $R^2$ values ($\hat{\beta} = .03, t=3.91, p<.0001$).

Q7: In Figure 3, would it be possible to include inter-human reliability

As the sequences of observations were randomly assigned and not identical, inter-rater reliability cannot be calculated.

Q8: Is it not clear how to control the number of training data in CLIP and SimCLR?

We'll add: "Comparing image-language pretraining with augmented image training remains challenging, as language may provide information beyond an augmented version of the same image."

Q9: the model's number of parameters might also be a confound

Model size is important (Figure 4B), but some smaller CLIP models outperform larger ones, and some large self-supervised models are outperformed by smaller CLIP models. Appendix Table 3 shows no relationship between size and score for CLIP models, suggesting performance isn't solely driven by model size.

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 alignment method improved predictive accuracy » (line 15).

We added more baseline comparisons and tested alignment methods on two additional datasets (Rebuttal Figure 1 & 2). DreamSim consistently improves alignment in its test set, gLocal shows the largest improvements in our task, while Harmonization degrades alignment. We are limited to the models trained by the original authors, which allows us to only offer a qualitative report. Harmonization alignment is only possible for supervised models by the nature of the method.

[1] SLIP: Self-supervision meets Language-Image Pre-training, Mu et al., arxiv 2021

[2] Learning strategies in amnesia, Speekenbrink et al., Neuroscience & Biobehavioral Reviews 2008

[3] A unifying probabilistic view of associative learning, Gershman, PLOS CB 2015

[4] Heuristics from bounded meta-learned inference, Binz et al., Psychological Review 2022

[5] Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Hebart et al., Nat. Hum. Beh. 2020

It would have been very insightful to include a comparison with generative models.

Thanks for the great suggestion! We now included 3 masked autoencoder models[1] trained on ImageNet in our comparison. We chose these specific models because they are SOTA classification models trained in a generative manner. They don’t do well in predicting human choice (mean $R^2$ = 0.05, min $R^2$=0.03, max $R^2$=0.06). Therefore, it appears that training generative models doesn't provide a particular advantage, at least in our comparisons. We will include these results in the camera-ready version.

M1: There is a potential bias in using the THINGS database for both training and evaluation, especially with the gLocal models,

Yes, this is an important limitation that we have also pointed out in line 279. Nevertheless, we wanted to include this comparison with that caveat. To our surprise, despite the shared data across the gLocal training and our task, we did not observe consistent improvements in model fit with the gLocal compared to the baselines.

Addressing this concern by diversifying the evaluation datasets (...) You could use a measure from DreamSim and Harmonization as control.

Thank you for these suggestions. We have now compared the alignment methods tested across two additional tasks that do not use the THINGS data (Rebuttal Figure 2). Similar to our initial findings, gLocal leads to improved alignment in limited cases. We found that Dreamsim consistently improves alignment on its test set, but no consistent improvements are observed on the independent dataset.

Additionally, we have incorporated four more baseline models for Harmonization and one more for DreamSim. These new baselines led to some modest improvements in our task, although they were not consistent and were smaller than those achieved by the gLocal models. As observed in our task, Harmonization models tend to reduce alignment in most cases. We will include these results in the camera-ready version.

We would also have liked to make this comparison for the Harmonization metric on the Harmonization dataset. However, this was not possible because the Harmonization metric requires a trained Imagenet classification head, which Dreamsim and gLocal models do not have.

(...) or providing a more detailed justification for using THINGS would strengthen the paper.

We chose to conduct our experiments with the THINGS data because the ground truth features that we used were only available for THINGS. This is because the ground truth features are a low dimensional embedding that was learned to predict human odd-one-out similarity judgements on the THINGS dataset. This embedding is central to the design of our tasks and unfortunately, it does not generalise to new image data. Therefore, we had to use the THINGS images. We will add this justification to the camera-ready version.

m1: (…) Standardizing all activations before performing regression could mitigate this issue and provide a fairer comparison.

Thanks for pointing out this important detail. This is indeed what we have done in our comparisons. We scaled the features using the mean and the standard deviation of the training data and applied the same transformation to the test data. We did this both in our mixed-effects models and also the linear and logistic regressions that were used to fit the features to the tasks.

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

Thanks for the suggestion. We have now made this comparison including the new models we added and observe that min-max scaling ($\rho$=-0.23, $p$=0.04) and standard scaling ($\rho$=-0.24, $p$=0.03) yield near identical results.

m3: Figure 7 lacks clarity on which models are used in the harmonization method?

We were only including the Harmonization models for which we had baselines. We have now added baselines for all the Harmonization models and therefore include all Harmonization models in our comparison. The results are highly similar to what was reported in the initial submission.

m4: Could you also add a metric to complement McFadden (e.g., avg accuracy) ?

Yes, please see Rebuttal Figure 3. The maximum accuracies that are achieved are 68% for the category task and 67% for the reward task. Sorting the models by accuracy and $R^2$ alters the ordering of the models slightly, although the rank correlations are very high both for category ($\rho$=.97) and the reward ($\rho$=.93) tasks and our qualitative results stay the same, where 9 of the top 10 models are CLIP variants in both tasks. We will include the average accuracy in the final version.

[1]He, K., Chen, X., Xie, S., Li, Y., Dollár, P., & Girshick, R. (2021). Masked Autoencoders Are Scalable Vision Learners.

The authors claim that their alignment metric is better than existing alignment metrics as it requires "generalisation and information integration across an extended horizon".

We believe that our method is not necessarily better than other metrics but that alignment is multifaceted. We will clarify this stance further in the main text by updating line 248 as follows: "Previous work has predominantly focused on simple image exposure and similarity judgments. We believe our findings complement this research by addressing unexplored aspects of alignment, which is generalisation and information integration across an extended horizon."

how correlated the new metric is to existing metrics

This is an excellent question that’s crucial for adequately placing our work in the field of alignment. To address this, we compared our alignment metric to four different metrics computed on different datasets. The results are shown in Rebuttal Figure 1. We will add the following to the manuscript:

"We compared alignment on four datasets against ours, which included:

I) Comparison of 22 models on the odd-one-out similarity judgments on THINGS [1] ($\rho=0.54$).

II) Comparison of 79 models on pairwise similarity judgments from an independent dataset [2] ($\rho=0.45$).

III) Comparison of 79 models on two-alternative forced choice data, which is a part of the NIGHTS dataset from the DreamSim paper that was not used to train the model [3] ($\rho=0.28$).

IV) Comparison of 25 models on alignment to the validation split of the ClickMe dataset [4,5], which was used to fine-tune the Harmonizer models ($\rho=-0.43$).

While our metric positively correlated with the metrics from the first three datasets, there were important differences. We found a negative correlation between our metric and the ClickMe-Harmonizer feature alignment, suggesting that pixel-level alignment and semantically bound global image alignment might be at odds."

For the odd-one-out data, we compared the models reported in the original paper [1]. For [2,3], we ran the comparisons for all the models. The ClickMe-Harmonizer alignment was only computed for supervised models, as the method requires computing gradients for ImageNet classes, which we could only do for the supervised models that had ImageNet classification heads.

the authors could take human pairwise similarity judgments and treat them as features from which to learn the task

This is a great suggestion. Unfortunately, pairwise similarity judgements do not exist for THINGS. However, the ground truth features of the task were generated from an embedding optimised to predict odd-one-out similarity judgements on THINGS[7]. We report how well these models predict human choice in Figure 2. While the ground truth model ranks in the top 13, it is surpassed by several multimodal models.

I didn't see an analysis of inter-rater reliability/noise roof.

The sequence of observations that were given to the participants in our tasks were randomly assigned and none were identical. Therefore it is not possible to calculate inter-rater reliability. However, the question of how good can alignment be is important. We initially thought the models trained on the ground truth features would establish a ceiling. However, to our surprise, several models exceeded this.

why were participants with less than 50% accuracy excluded?

In online studies, it is common to have noisy data, where sometimes participants don’t engage with the task faithfully. Therefore, it is common to exclude participants who perform below chance. Nevertheless, when we fit our models to all participants’s data (Rebuttal Figure 4), the $R^2$ values we obtain are highly correlated with the $R^2$ values after exclusion ($\rho=0.96$ for category learning, $\rho=0.66$ for reward). Our qualitative results stay the same, where the top-ranking models are CLIP variants. Therefore, our findings are not dependent on this filtering step.

why might the alignment numbers be so low?

We believe there are several reasons for this. First, it is difficult to predict human choices early in the task for any model, given the noise associated with uncertainty. Indeed, when we look at $R^2$ values in the first half of the task, the maximum values are 0.05 and 0.1 for the category and the reward tasks. Whereas if we look at the second half, they go as high as 0.22 and 0.21, which is in the range that you suggested. Regardless of the split, the top-performing models are CLIP models. This is a unique difficulty of learning tasks, as unsupervised similarity judgements do not require learning and therefore a similar noise effect is not expected early in the task.

Second, even when we use the ground truth features to predict human choice, we see that $R^2$ values aren’t high, which again speaks to the difficulty of the task.

Lastly, it’s difficult to compare McFadden’s $R^2$ values across tasks. Some simple learning paradigms might have best-fitting models with $R^2$ values <.2[7], whereas other similar paradigms can go up to the range you suggested [8].

[1] Human alignment of neural network representations. Muttenthaler et al., ICLR 2023

[2] Evaluating (and Improving) the Correspondence Between Deep Neural Networks and Human Representations. Peterson et al., Cognitive Science 2018

[3] DreamSim: Learning New Dimensions of Human Visual Similarity using Synthetic Data. Fu et al., NeurIPS 2023

[4] Harmonizing the object recognition strategies of deep neural networks with humans. Fel et al., NeurIPS 2022

[5] Learning what and where to attend, Linsley et al., ICLR 2019

[6] Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Hebart et al., Nat. Hum. Beh. 2020

[7] Finding structure in multi-armed bandits. Schulz et al., Cognitive Psychology 2020

[8] Model-Based Influences on Humans’ Choice and Striatal Prediction Errors, Daw et al., Neuron 2011

Did you consider fitting a mapping between the model features and participant responses more directly? If not, why not?

We agree with the reviewer that fitting the linear/logistic regression models directly to human choices is an interesting thought (as it has for example recently been applied in a different domain by [1]). This modelling approach typically requires a lot of data to work well. For example, [1] has used data from around 1 million human choices, whereas we have only 60 or 120 choices per participant. Hence, we expected that this approach would not work well in our setting. Nevertheless, we have now also tried to fit the embeddings directly onto human choice. However, we received much poorer fits (mean $R^2$= -0.76, max $R^2$= -0.05, min $R^2$= -3.98), as this approach does not directly capture the structure of the task and likely leads to overfitting.

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.

Under the Behavioural Analyses section in Appendix A, we will add the following in the camera-ready version:

“We used mixed-effects logistic models for both category and reward learning analyses. For category learning, we predicted correct responses per trial, using trial number as a fixed effect and including participant-specific random effects for intercept, trial number, and assigned task rule. In the reward learning model, we predicted whether the image on the right is selected, incorporating the trial number, reward difference between images, and their interaction as fixed effects. These factors, along with the assigned task rule, were also modelled as participant-specific random effects. Both models effectively captured task structure, learning progression, and individual variability in performance.”

And under the Modelling in Appendix A, we will add the following:

“We used mixed effects logistic regression models with leave-one-out predictions to assess participant choices in both tasks. For category learning, we used logistic regression probability estimates as predictors. In the reward learning task, we used the difference in estimated rewards from linear regression models as predictors. In both cases, these predictors were included as both fixed and random effects, allowing us to account for individual differences while maintaining the group effects. These correspond to the following models in R formula notation:

Category Learning:

human_choice ~ -1 + probability_estimate + (-1 + probability_estimate | participant)

Reward Learning

human_choice ~ -1 + estimated_reward_difference + (-1 + estimated_reward_difference | participant)

Where -1 denotes no intercept."

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.

Thanks for raising this point. It has been shown that using “fun” cover stories improves data quality in learning tasks[2]. In our experience, this also makes the tasks more engaging, especially in online settings. Furthermore, before starting the main experiment, participants were required to answer a few comprehension questions, and they were not allowed to participate in the main task until answering these questions correctly. Therefore, we believe participants had a good understanding of the task. We will include these questions in the Appendix for the camera-ready version.

Typos "Exemplary learning curves" (pg 4)

Thanks. We will change this to “Example learning curves”

[1] Peterson, J. C., Bourgin, D. D., Agrawal, M., Reichman, D., & Griffiths, T. L. (2021). Using large-scale experiments and machine learning to discover theories of human decision-making. Science, 372(6547), 1209-1214.

[2] Feher da Silva, C., Lombardi, G., Edelson, M., & Hare, T. A. (2023). Rethinking model-based and model-free influences on mental effort and striatal prediction errors. Nat. Hum. Behav., 7(6), 956–969.

Thank you for engaging with the review process and raising the score. We are happy we could clarify our approach and the motivation behind it.

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.