acceptodds
Withdrawn from ICLR 2027

Beyond Prediction Accuracy: Target-Space Recovery Profiles for Evaluating Model–Brain Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Alignment between vision models and the human brain is often evaluated by how accurately model representations predict brain responses. When two models achieve similar prediction accuracy, this measure does not establish whether they recover the same brain-response structure. We propose a framework for comparing models through their predictive subspaces in the target subject's response space. For each target subject and visual region, we construct a weighted reference from functional magnetic resonance imaging (fMRI) responses across multiple trials of the same images. In the main analysis, this reference averages predictive-subspace projectors obtained by ridge-regularized reduced-rank regression between trial-group means. A recovery profile measures each model's subspace overlap with the weighted reference directions. Differences between profiles obtained from other subjects' responses, matched for prediction accuracy within each target, provide a brain-to-brain benchmark for model differences. We analyzed responses to 515 shared natural images from eight Natural Scenes Dataset (NSD) subjects in early and intermediate visual areas (V1, V2, V3, and hV4). We compared ImageNet-pretrained and randomly initialized models of the same architecture, matched within 0.01 in held-out prediction accuracy. Pretrained models had higher mean recovery-profile scores in 25 of 29 pairs. The mean difference remained positive with an alternative reference based on signal covariance estimated by Generative Modeling of Signal and Noise (GSN). These results show that recovery profiles can reveal differences associated with pretraining, even when prediction accuracy is closely matched.

Primary area:
applications to neuroscience & cognitive science
Keywords:
NeuroAI, model-brain alignment, human visual cortex, fMRI, Natural Scenes Dataset, encoding models, vision models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vtVpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes “recovery profiles.” These measure how much of a reliable brain-response subspace a model’s predictions cover. It builds the reference from repeated fMRI trials and compares models against brain-to-brain differences. On NSD early visual areas, pretrained models score higher than random models with the same accuracy in 25 of 29 pairs.

Critical strengths
  • The problem is interesting. Two models with the same accuracy may not capture the same brain signal.
  • Using other subjects as a benchmark gives a natural scale for model differences.
  • There are good robustness checks (e.g. NSD-synthetic, etc.)
Critical weaknesses

The main test compares only pretrained and random networks. Several previous papers already show that pretraining improves model–brain alignment, so this is an easy contrast. In higher visual areas, the method does not seem to separate trained models beyond brain-to-brain variability. In early visual areas, differences are larger but lack significance tests. These early regions are also tricky: alignment there may reflect low-level stimulus features rather than actual neural processing.

Questions for authors
  • Can you test the method on harder contrasts, such as trained models with different architectures or objectives?
  • In higher visual areas, the method does not seem to separate trained models beyond brain-to-brain variability. Would more images or repeats fix this, or is the method limited there?
  • In early visual areas, can you add significance tests for differences between trained models? And can you show that alignment there reflects neural processing rather than low-level stimulus features?
Reviewer_rV5kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a method for comparing brain-predictive models beyond their held-out prediction accuracy, based on the subspaces of brain responses they predict. Using ridge-regularized reduced-rank regression (RRR) to map between different trial splits within the same subject, the method first constructs a target reference space of directions that are preserved across trials. For any source, such as the fMRI responses of other subjects or the features of a vision model, a predictive subspace is then constructed in target coordinates. The source is evaluated by how well this subspace recovers the reference space, in addition to its prediction accuracy. Differences between other subjects’ brains at matched accuracy serve as a benchmark for how large a model difference must be to count as meaningful. The authors evaluate a range of pretrained and randomly initialized models of different architectures and find that models matched in prediction accuracy can differ substantially in how well they recover the reference space.

In my reading, the paper contributes a relevant additional dimension for model-brain benchmarks with a simple and elegant method. However, its novelty relative to CCA and related subspace comparison methods is not clearly established, and it remains unclear how differences in recovery should be interpreted. The results sections are also difficult to follow without the appendix.

Critical strengths

Relevant question and an elegant method: Prediction accuracy alone often does not distinguish between brain-predictive models, and the paper offers a methodologically simple and clean way to compare them along an additional dimension. It integrates very recent literature, and the core method in Section 3 is described rather clearly. I would consider using the method myself.

Brain-to-brain variability as a benchmark: Using differences between accuracy-matched brain sources as a reference for model differences provides a principled scale for deciding whether two models differ meaningfully. The evaluation covers two reference estimators as well as pretrained and randomly initialized models.

Critical weaknesses

Lack of self-containedness: Sections 4 and 5 in particular are hard to follow. They contain many subheadings, several of which are not intelligible on their own, and the results lack connecting text and guidance for the reader. Section 3 is clearer but also relies heavily on the appendix for key details, such as how the reference hyperparameters are selected and why cvPCA is used only for spectral comparisons. This makes it difficult to assess the results and how they support the paper’s conclusions. If the paper becomes readable after revision, I would raise my score.

Unclear interpretation and novelty: It remains unclear what an overlap, or lack thereof, between the predictive subspaces of two sources means. Coverage is purely geometric: as the authors note, even weakly predictive directions can increase it, and a source with a higher selected rank spans more of target space. Differences in coverage between accuracy-matched models may therefore partly reflect the rank chosen by the one-standard-error rule rather than a difference in alignment with the brain. In addition, the relation to CCA and to existing subspace and similarity measures is not clearly discussed, so it is unclear what the framework captures that these do not. This is critical because the main finding, that accuracy-matched models differ in recovery, is only informative if recovery can be interpreted and adds information beyond existing measures.

Questions for authors

How does the framework relate to CCA-based comparisons? Directional coverage corresponds to squared cosines of principal angles, and the weighted profile resembles projection-weighted CCA. What does the RRR-based construction capture that a regularized CCA between source and target, or existing similarity measures such as CKA or RSA, would not? Do those measures already distinguish the accuracy-matched models?

How should a difference in coverage between two accuracy-matched sources be interpreted? Is a model with higher coverage better aligned with the brain, or could it span more reliable directions while predicting each of them weakly?

Reviewer_AzD7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims at answering an interesting problem, which is that encoding accuracy alone cannot tell whether two models capture the same aspects of brain responses, and proposes "recovery profiles" as a complementary evaluation. For each subject and brain region, the method first identifies the few response patterns that are reproducible across repeated presentations of the same images via a split-half analysis. It then measures how many of these patterns each source (either a deep neural network or another subject's brain responses) can reproduce through a linear mapping. Using NSD data from early visual cortex (V1–hV4), the authors show that ImageNet-pretrained networks reproduce more of these patterns than randomly initialised networks with matched prediction accuracy, and support this with a series of robustness analyses.

Critical strengths

Main strengths of the paper:

  1. This is an important question and the method is model-agnostic: Prediction accuracy is the standard way to compare models with the brain, but it cannot tell whether two equally accurate models capture the same response patterns. The proposed score addresses this directly: the reliable patterns are defined from the target subject's repeated trials only, every source (a deep network or another subject) is scored against the same patterns, and differences are calibrated by the variability between equally accurate subjects. The example in Fig. 3A illustrates the point well: pretrained and random VGG-16 have the same accuracy (r ≈ 0.31) but very different scores (0.74 vs 0.42).
  2. The main result is well tested. Pretrained networks outperform random ones in 25 of the 29 accuracy-matched pairs, and the authors show that this holds up under many checks: a different way of estimating the reliable patterns (GSN), patterns estimated on separate images, head-motion correction, several random seeds, different analysis settings, and a different stimulus set (NSD-synthetic). They are also open about the limitations of their approach, which makes the result convincing within the setting they study.
Critical weaknesses

Main weaknesses are:

  1. At the moment it is unclear what the score adds beyond prediction accuracy: The score checks whether a model can produce a response pattern at all, not whether it predicts it well (the authors note this in Sec. 3.2). As a result, the score can increase simply because a model's linear mapping uses more dimensions, even when its prediction accuracy does not improve, and matching models on accuracy does not control for this (Sec. 6).
  2. The scope is too narrow to show that the method is useful: The main analysis covers only early visual cortex (V1–hV4), a single contrast (pretrained vs random networks), and a reference estimated from only 103 images × 3 trials per fold (Sec. 4). In higher-level regions, where comparing trained models matters most, the authors report smaller model differences relative to brain-to-brain variability (Sec. 6).
  3. The method is partly unclear and also hard to follow: The main text does not explain how model features are extracted. Section 3 introduces many components (the trial-split reference, predictive subspaces, coverage, recovery profiles, ProfileMean, effective rank and the brain-to-brain threshold) in a dense way, with frequent pointers to the appendix, and Figure 1 does not make the steps easy to follow; it takes several readings to understand how the reference patterns are actually obtained. The results also depend on several settings: the rank cap (20) and the one-standard-error rule for choosing the rank, which the authors show affect how concentrated the reference is (Sec. 5.1); the number of directions K = 10; the 0.01 accuracy-matching tolerance; and the 90th-percentile brain-to-brain threshold. Some sensitivity checks are reported (Sec. 5.3), but there is little guidance on how to choose these settings, which matters for a metric that others are meant to adopt.
Questions for authors

Some questions and clarifications

  1. In the 29 matched pairs, which ranks are selected for pretrained vs random models, and does the effect hold at a fixed rank?
  2. How much of ProfileMean comes from the first direction alone?

Some suggestions 3. Validate the score on simulated data where the answer is known: create two artificial sources that predict the brain equally well but capture different response patterns, and show that the score tells them apart while prediction accuracy does not. 4. For comparisons between models only, which do not require other subjects, use each subject's own NSD images (thousands, each shown three times) instead of the 515 shared ones. This would give more reliable reference patterns and could show whether the score separates models in higher-level regions. 5. For future work, it would be interesting to apply recovery profiles to language, comparing large language models with language-selective brain areas, where many models reach similar encoding accuracy and a finer-grained comparison would be especially useful. This would require fMRI datasets with repeated presentations of the same language stimuli (e.g., naturalistic story-listening datasets with a repeated test story for instance "A natural language fMRI dataset for voxelwise encoding models"), and could test whether the score separates models (e.g., trained vs untrained, or models of different sizes) beyond what prediction accuracy shows.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.