acceptodds
Withdrawn from ICLR 2027

Examining the Temporal Dynamics of Human Alignment by Probing Model Learning Windows

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Critical periods in biological development are temporal windows when experiences leave lasting effects on a system. Timing of critical periods differs based on the function being learned, with different sensory, cognitive, and linguistic systems following distinct schedules. In deep neural networks, critical periods have been identified primarily through task accuracy, collapsing this diversity into a single curve. We ask whether critical periods differ when measured through representational alignment with the human visual cortex. We train ResNet, ConvNeXt, and ViT under supervised ImageNet classification, ViT under contrastive language-image pretraining (CLIP), and pretrained CLIP-ViT under a fine-tuning task, injecting transient perturbations to inputs and targets at different training stages. We measure the effects at three levels: immediate change in internal representations (via similarity to human behaviorally-derived embeddings and fMRI activations), the trajectory representations follow as normal training resumes, and generalizability of learned embeddings via linear probes on out-of-distribution tasks. Early input-side perturbations cause the most disruption to representational structure, and models that recover task performance also recover human alignment. In supervised classifiers, human alignment and task performance recover from all perturbations. In CLIP, early and mid perturbations permanently damage task accuracy, yet after mid-training perturbations, brain alignment recovers in every case and behavioral alignment in two of three, showing the two objectives can dissociate. Late input-side perturbations have a different effect on ImageNet classifiers: representations of within-category objects are pulled closer together, distinguishing between-category boundaries and transiently improving human alignment. Critical periods in neural networks depend on what is measured, not just when. If late perturbations push representations toward human-like organization with little harm to performance, targeted interventions could steer representational alignment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Human Alignment, Critical Periods, Vision Models, Training Dynamics, Representational Similarity Analysis

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 4 reviews, above 27% of the others.
Reviewer_YY3zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This study uses single epoch perturbations to assess temporal dynamics of human representational alignments during training stages. The authors claim that task classification accuracy can dissociate from behavioral and neural alignment, tested supervised models recover in accuracy and behavioral and neural alignment after all perturbations while CLIP VIT_B/32 does not recover in task accuracy (after early and mid perturbations), and late perturbations can transiently boost alignment.

Critical strengths
  1. This study tests both neural and behavioral alignments of the model with humans under the training-phase perturbations which can help increasing our understanding of how models form human like representations in learning stages both in brain activity and perceptual judgments.

  2. The category/class-level view is an interesting useful perspective to take.

Critical weaknesses
  1. Some experiments need important controls. Since CLIP has been perturbed for a larger percentage of the total training time than other models and the early perturbation is during the learning rate warm-up. The authors’ arguments against LR concern is not sufficient considering that the models also differ in data and schedule, other controls listed below are needed to rule these factors out.
  2. The alignment boost disappears after one epoch and is only present in 3 single seed runs, so the promise that the targeted interventions could steer alignment is under investigated and the proposed mechanism is only supported correlationally.
  3. Considering that only CLIP VIT-B/32 cannot fully recover from the early and mid perturbations, making a general claims (in abstract) could be considered overstatements that have only been addressed in the limitation section.
  4. Some main claims are based on results of a single seed per from-scratch model.
  5. No confidence intervals reported for per-run alignment results of the from-scratch models.
Questions for authors

Further experiments that could answer the questions below can address my concerns:

  1. How would your results change if the ImageNet models are perturbed for 3% training time (e.g. 3 epochs)?
  2. Would other onset perturbations (between epoch 13 to 28) in CLIP show permanent accuracy/alignment drops? This gives a better understanding of where the window closes.
  3. Have you considered testing your hypothesis on other models such as DINO/SimCLR to compare the results with CLIP?
  4. How do you explain that the CLIP VIT-B/32 results (mid-onset +0.004 to +0.013 for neural and −0.016 to +0.015 for behavioral alignment) are described as above baseline when they are comparable to baseline median epoch to epoch changes (in behavioral alignment) reported in the appendix?
Reviewer_xsiYOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

Prior work on critical periods in deep networks has measured the effect of early disruptions almost entirely through task performance. This paper asks a different question: do those disruptions also change how a model represents the world, and in particular whether its internal organization ends up human-like? The motivation is reasonable. Models with more human-like representations tend to be more robust and to generalize better, so anyone designing a training procedure would want to know when human-likeness is most fragile, and whether that window coincides with the one for accuracy. To test this, the authors corrupt training for one epoch at an early, mid, or late point. They use two methods of perturbation:

  • Injecting Gaussian noise or uniform images on the input side.
  • Shuffling the ground truth labels on the target side.

They then compare how each perturbation, at each onset, affects task accuracy and the alignment of the model’s representations with human behavioral and neural data.

Critical strengths

The perturbations themselves are not new. The novelty lies in connecting critical-period analysis to human representational alignment, and I think that connection is worthwhile. However, the broader observation that ImageNet accuracy and human-likeness don’t go hand in hand has been reported before.

Critical weaknesses
  1. The role of the learning rate is not disentangled. The learning rate and its schedule are among the main factors governing how much a network can still change, yet the paper doesn’t manipulate either. Its discussion of the learning rate is limited. This matters because onset timing and learning-rate magnitude are confounded in the design. Early perturbations land near the peak of the schedule, and late ones arrive after it has decayed to small values. As a result, the experiments can’t separate “when in training” from “how large the learning rate was.” A reader could reasonably ask whether a critical window is simply a property of the schedule. A large early learning rate allows big moves, and as it decays the model settles and becomes harder to change. A scheduler clearly contributes to a closing window. If the authors believe learning rate schedulers do not play a huge role in studying critical windows, then they should clarify why this is not the case

  2. Half of the class-versus-category result is close to built-in. A classifier needs separated classes in order to classify, so the link between class gap and accuracy is largely expected. The informative half of the result is the category gap predicting alignment, and the paper should weight its claims accordingly.

  3. Related work: biological framing. Critical windows differ across species as well as across functions, and the related work could make that point more explicitly. I’d suggest adding imprinting in chicks, which is one of the clearest and most widely known examples of a critical window and fits naturally alongside the kitten, barn owl and sparrow examples already discussed.

Questions for authors
  1. Can the authors include a constant-learning-rate control, or otherwise show that the onset effects aren’t explained by learning-rate magnitude at each onset? Additional experiments are needed to narrow down the results to the claims presented in the paper.
  2. How was the category-gap threshold chosen, and do the results hold without it?
  3. Does the category-gap account hold under a different category hierarchy?
Reviewer_RPuCOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to evaluate 'critical periods' in computer vision models, by measuring task accuracy and brain/behavior alignement during perturbed training. They show that supervised models recover accuracy and alignment and accuracy under all perturbation, while CLIP accuracy can stay impaired even if alignment recovers.

Critical strengths

I am not an expert in this field but believe this particular test, especially in relation to behaviour and brain alignment, is new.

Critical weaknesses

Methods:

  • I am wondering whether this approach wouldn't be more stable by first doing systematic comparisons to pretrained networks (e.g. Dino V3). The alignment would not be limited to a small number of images and categories, as it is currently done, and could allow identifying a more systematic set of phenomena elicited by early, mid and late perturbation, which could then be checked with brain/behavior.

Contribution:

  • The notion of critical period generally relates to a specific time window during which a feature can be specifically learnt. Here the authors evaluate the perturbation effect on 3 preselected phases of training. I would have expected, from the goal stated, a much more thorough investigation of the training periods that may exists for learning specific features.
  • The authors do show an interesting dissociation between coarse category structure vs fine class structure. However, the features effectively lost or preserved remain unclear.

Impact: -These analyses are new means to evaluate the stability (or training trajectory) of computer vision models under perturbation. However, I am not convinced we evidence, and thus learnt a new principle about critical periods.

Questions for authors

Why is alignment not first evaluated against pretrained representations from SOTA models? Do the early, mid, and late perturbation effects replicate against these models? Why were only three perturbation checkpoints tested, rather than something more systematic? What new principle about critical periods should we take from the results; do they inform critical periods in human learning?

Reviewer_KiexOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates how perturbations at different stages of neural network training affect human representational alignment. The authors perturb model inputs or targets for one epoch and track changes in task accuracy, behavioral alignment, brain alignment, and downstream transfer. They find that accuracy and human alignment can recover differently, particularly in CLIP. They also suggest that accuracy depends more on fine-grained class structure, while human alignment is more related to broader category structure. Some late-stage perturbations temporarily improve alignment.

Critical strengths

I think studying how human alignment responds to perturbations throughout training is an interesting perspective. The intervention-based approach provides insights beyond simply tracking alignment during normal training. The work also explores both alignment in brains and behaviors, and use CKA metric and NOD datasets as complementary ways to justify the evaluation.

Critical weaknesses

My main concern is that the paper is largely based on observation and is discriptive. The authors examine how perturbations at different training stages affect human representational alignment and task performance, but I am not convinced that the findings provide a sufficiently strong scientific contribution. The experiments show some interesting patterns, but they do not clearly explain why these effects occur or what new understanding we gain about the development of human-like representations.

The authors acknowledge several limitations, including using only one training seed for each of the four from-scratch models and different training recipes across architectures. I consider these fundamental weaknesses rather than minor limitations. In particular, the persistent perturbation effects are mainly observed in a single CLIP model, while all three supervised classifiers recover rapidly. Given the differences in training objectives, datasets, optimization schedules, and relative perturbation durations, it is difficult to determine whether these effects reflect critical learning periods or are specific to a particular training configuration. I think the current experimental evidence is not sufficient to support the broader conclusions about critical periods in human representational alignment.

I also feel that the results are still preliminary, and more experiments are needed to establish their broader significance. For example, how could these findings help us develop better training strategies to improve human–AI alignment? The authors show that some late-stage perturbations temporarily improve alignment, but these improvements disappear shortly after normal training resumes. This is an interesting observation. More work can be done to show a clearer connection between the observed training dynamics and how they might inform future alignment methods.

Overall, I think the paper explores an interesting question, but the current scope and experimental evidence do not meet the bar for ICLR. The work is still at a preliminary stage apparently and would benefit from further experiments and a clearer scientific message. It may be more suitable for a workshop for authors to receive feedback and further develop the work.

Beyond these scientific concerns, I have a few issues with reproducibility and presentation. The anonymous code repository linked in the paper was inaccessible when I checked it. The AI Use Statement is also quite general. Given that the authors report using AI tools for methodology, experiments, dataset cleaning, code, and writing, I think they should specify which tools were used and clarify their contributions. Finally, the AI Use Statement is numbered as Section 6, while the Reproducibility Statement is unnumbered. The formatting should be consistent.

Questions for authors
  • See Weaknesses section
  • What new scientific insight does the perturbation framework provide beyond the baseline training dynamics? The paper already shows that human alignment can increase or decrease during normal training. What do the intervention experiments reveal about the underlying learning process that cannot be inferred from these baseline trajectories?
  • How robust are the critical-period findings to random seeds and training configurations? Since lasting deficits are observed primarily in one CLIP model, could the authors provide stronger evidence that these effects are attributable to perturbation timing rather than differences in the training objective, optimization schedule, or perturbation duration?
  • How much of the class-versus-category explanation is driven by the evaluation metrics? Could the authors examine whether the relationship between category separability and human alignment remains after controlling for coarse semantic categories, for example through within-category RSA?
  • How could the findings be used to improve human alignment? The reported late-stage improvements are temporary. Can the authors identify a concrete intervention strategy, supported by their findings, that produces sustained improvements in human alignment without substantially reducing task performance? Alternatively, what specific prediction about such a strategy follows from the observed dynamics?

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

Related papers

Discussion (0)

Sign in to comment.