acceptodds
Under review as a conference paper at ICLR 2027

Early Structural Experience Shapes Visual Representation Learning

Abstract

During development, infants initially experience a visually degraded world, with limited spatial acuity, contrast sensitivity, chromatic sensitivity, and binocular depth perception that gradually mature over the first months of life. This developmental trajectory may bias early learning toward more stable structural cues, such as boundaries and shape, before richer appearance cues become available. Early-learned structural representations could subsequently serve as attractors that guide the interpretation of richer visual experiences, consistent with infants’ and children’s preference for line drawings and cartoons. Here, we investigate this developmental hypothesis in deep neural networks by using semantically meaningful line drawings as a surrogate of early structural visual experience. Line drawings preserve object boundaries, global shape, and part configurations while suppressing much of the texture and color information present in photographs. We find that early exposure to line drawings systematically reshapes subsequent learning from photographs toward more shape-centered representations. These representational changes improve performance on tasks requiring geometric and spatial understanding, including segmentation and depth estimation, in addition to enhancing recognition, detection, data efficiency, and knowledge distillation. Our findings suggest that early learning based on stable structural information creates a representational foundation to guide the learning of richer visual knowledge in later experience.

Primary area:
applications to neuroscience & cognitive science
Keywords:
visual representation learning, structural inductive bias, shape bias, line drawings, curriculum learning, transfer learning, data efficiency, computer vision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Mb2ROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work investigates how infants' early degraded vision shapes visual representation learning. To simulate degraded vision, the authors generate line drawings from photographs. Then, they train vision models while shifting the training data from line drawings to photographs. Controlled experiments show that the curriculum improves the learning efficiency and the shape bias of the models.

Critical strengths

Evaluations are extensive and appropriate.

Critical weaknesses

This work is not mature enough for presentation in this conference.

  1. The paper does not cite major studies that directly relate to the present research questions [1,2].
  2. The paper presents methodological issues. First, measuring the cosine similarity between CLS tokens does not reflect changes in the representational geometry. If all representations of layer l are rotated by the same angle at layer l+1, the cosine similarity between representations at layers l and l+1 would be small. Yet the relationships between l+1 representations are identical. To address this limitation, I recommend to perform a representational similarity analysis (RSA). Second, the accuracy on Imagenet-100 is relatively low, compared to results obtained with classical SSL methods (e.g. [3]). Third, controlled experiments are required to verify that the performance gains do not stem from increasing the number of training images: Photo -> Line and Photo + Line.
  3. In cognitive science, the link between the line drawings and the early degraded vision is unclear. One major property of early degraded vision is low contrast sensitivity. Thus, low spatial frequencies (such as thin edges) should be difficult to detect, which is not the case there. In machine learning, the experiments do not meet the standards of the field, which is Imagenet-1k training.
  4. Some paragraphs are hard to follow, such as "Concentration of channel responses".

[1] Cai, Y., Lin, Q., Nunna, B. S., & Zhang, M. (2026). Learning to See Through a Baby's Eyes: Early Visual Diets Enable Robust Visual Intelligence in Humans and Machines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 13766-13780). [2] Lu, Z., Thorat, S., Cichy, R. M., & Kietzmann, T. C. (2026). Adopting a human developmental visual diet yields robust and shape-based AI vision. Nature Machine Intelligence, 1-14. [3] Aubret, A., Teulière, C., & Triesch, J. (2026). Seeing the whole in the parts with self-supervised representation learning. Neurocomputing, 132973. desvribe the method in the related works

Questions for authors

I will increase my score to acceptance if the authors:

  • clarify how the present results relate to [1,2], especially [2] regarding the shape bias.
  • clarify that the improvements come from the proposed curriculum, and not from using more images.
  • use RSA to compare representations across layers, or remove the claims related how representations change through the neural hierarchy.
  • explain the precise relationship between the photographs and early degraded vision.
  • test the model on Imagenet-1k.

If that is not possible, I encourage the authors to pursue their research direction, re-frame the story, and resubmit the paper if the preliminary results are confirmed.

3 comments

We thank the reviewer for the detailed feedback and for recognizing the breadth of our evaluations.

1. Related work and shape bias

We now discuss Cai et al. and Lu et al. in “Developmentally Motivated Visual Learning” and compare adaptations of their sensory transformations within our supervised STL10 setup. Both sensory and structural curricula increase shape bias. Line→Photo achieves the highest recognition accuracy among the tested conditions, while blurred grayscale yields the highest shape bias; see “Developmental sensory curricula.” Our contribution examines the broader representational and transfer consequences of early structural learning.

2. Training budget and controls

The recognition and mixing controls use matched total training budgets, counting all stages. Drawings come from the same underlying photographs.

Photo+Line achieves the highest recognition accuracy (75.19%), while Line→Photo produces the strongest shape bias (61.45%). Both outperform Photo→Mixed (69.09% recognition, 38.18% shape bias) and have higher mean neuronal structure sensitivity at all four stages. The overarching response provides the full comparison. Photo→Line is reported in “Recognition across training and evaluation modalities.”

3. Representational geometry

We replaced cross-layer CLS cosine alignment with RSA in the ViT. Using the same evaluation images, we construct cosine-distance dissimilarity matrices and compute Spearman correlations between each layer and its model’s own final layer. These correlations are higher for Line→Photo at every intermediate layer; see “Representational geometry across network depth.”

4. Developmental interpretation

Line drawings serve as a computational surrogate for early structural representations, rather than infant retinal images. The pretrained generator supplies the structural prior; we do not assume it is innate in infants. Our hypothesis concerns how early structural experience guides later learning. Blurred-line controls retain recognition gains at all three blur levels, showing that the benefits persist when contour precision is reduced; see “Contour precision.”

5. ImageNet-1K and scaling

On ImageNet-1K, Line→Photo improves recognition accuracy with a YOLO-v8 backbone from 67.87% to 68.82%. Our ViT-B/16 experiments on ImageNet-21K provide further evidence under matched optimizer-step budgets. Together, these results show that the benefits of early structural learning extend to larger datasets and models.

6. Self-supervised accuracy

Our DINO and SwAV runs use ResNet-18 and standard solo-learn implementations. Aubret et al. use CO-SSL objectives and ResNet-50/RF-ResNet-50 backbones, so their absolute accuracies are not a matched comparison.

7. Writing clarity

We rewrote “Concentration of channel responses” to explain the averaging, ranking, and normalization explicitly, and distinguish this measurement from PCA dimensionality and RSA.

We hope these revisions address your concerns about the interpretation and soundness of the study.

I thank the authors for the response, which addressed some of my concerns.

Self-supervised accuracy: that is true, however different studies have also tested a Resnet18 and obtained a better performance than what is obtained in the present paper [1]. The exact SSL method do not exactly match, but they tested methods that are expected worse than DINO, which is presented in the present paper. Could the authors provide the results of their hyper-parameter analysis (especially, learning rates) during pre-training ?

ImageNet-1K and scaling: the authors report only the training loss of the ViT. Performance on a validation/transfer dataset must be reported. Focusing on a YOLOv8 backbone is also unusual; standard practice is to train a RN50 or ViT-S/B on Imagenet-1K. Comparison against baselines (Table 5), and most of the experiments, should be performed on Imagenet-1k. STL10 contains only 10 classes, there is a plethora of visual shortcuts (e.g. water = boat, sky=airplane) than can be learned to perform the classification task. Thus, performance on this dataset is not informative of the quality of a learning model, in general. If computational resources are a limitation, I encourage the authors to run the most important experiments on Imagenet-1k, and eventually do analyses on Imagenet-100, or at least Tiny-Imagenet/CIFAR100.

[1] Ozsoy, S., Hamdan, S., Arik, S., Yuret, D., & Erdogan, A. (2022). Self-supervised learning with an information maximization criterion. Advances in Neural Information Processing Systems, 35, 35240-35253.

Thank you for your follow-up and for recognizing the concerns addressed by our revisions.

1. Self-supervised baselines

We did not conduct an independent hyperparameter sweep. We used solo-learn and consistently reproduced its published ImageNet-100 results for ResNet-18 with 400 pretraining epochs: 74.92% for DINO and 74.28% for SwAV under offline linear evaluation.

Table 1 in Ozsoy et al. (2022) also draws its ImageNet-100 ResNet-18 baseline results from solo-learn. Its BYOL (80.32%), MoCo-V2 (79.28%), and VICReg (79.40%) results match those in solo-learn’s evaluation table, which also reports the DINO and SwAV baselines reproduced in our study:

https://github.com/vturrisi/solo-learn#imagenet-100

Although reproducing these results does not establish that the configurations are optimal, we are confident that our Photo-only runs match the published baselines. Under matched training budgets, Line→Photo improves DINO from 74.92% to 77.22% and SwAV from 74.28% to 77.62%.

2. ImageNet-1K and evaluation scale

We agree that repeating the principal curriculum comparisons on ImageNet-1K with ResNet-50 or ViT would strengthen the paper. Unfortunately, we lack the computational resources and time to complete these experiments within the rebuttal period. We also agree that the ImageNet-21K training-loss result alone does not establish improved validation or transfer performance.

Our evaluations beyond STL10 include supervised recognition on ImageNet-1K with a YOLOv8 backbone, where accuracy improves from 67.87% to 68.82%. On ImageNet-100, we evaluate supervised recognition, blur controls, and DINO and SwAV pretraining. We also evaluate object detection on PASCAL VOC and transfer to depth estimation on NYU Depth V2 and semantic segmentation on ADE20K.

These experiments provide evidence across datasets and tasks, but we recognize that the lack of standard-backbone ImageNet-1K comparisons is a limitation of the current study.

We hope the revisions and additional experiments have addressed your other concerns about the study’s framing, controls, and representational analyses. Thank you for your continued consideration. We would be happy to clarify any remaining questions.

Reviewer_ocVpOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors consider whether pre-training on line drawings improves visual representation learning. They use a pretrained model to generate line drawings from photographs, and train models in a stage-like manner of first, line drawings, and second, full photographs. They examine both supervised and self-supervised models across a range of benchmarks.

Critical strengths
  • Strong results from their method
  • I appreciate the thoroughness of the evaluations, which span object recognition, examining sample complexity, the shape bias, 3D shape perception, depth estimation
Critical weaknesses
  • The developmental motivation is a stretch. There is prior work on training networks with a "developmental visual diet" that is much more developmentally realistic than the current proposal in terms of matching infant acuity and contrast/chromatic sensitivity in different developmental stages, which should be cited and compared to the current paper https://www.nature.com/articles/s42256-026-01228-6
  • The current developmental framing is weak. Given that the current work focuses on line drawings, and there is no developmental visual stage that has children seeing the world like line drawings, I would insist they scrap the framing of their work as a "developmental hypothesis."
Questions for authors
  • The exact training procedure, including how the model transitions from line drawings to photos, is not clearly specified in the main text.
  • Were all the experiments compute-matched between the Line-Photo and Photo only conditions?
  • Why is the knowledge transfer experiment included? Do you have an explanations for these results? As it stands, I find it confusing and I'm not sure what it adds to the paper.
3 comments

We thank the reviewer for recognizing the strength and breadth of the results. We have clarified the developmental motivation, training protocol, and purpose of the knowledge-transfer experiment.

1. Developmental framing

We agree that infants do not experience a line-drawing stage. We now frame the study as a computational investigation of early structural learning, motivated by development, rather than a model of infant visual maturation. Line drawings provide a simplified surrogate for structural representations that may guide subsequent learning. The pretrained generator supplies this prior; we do not assume that infants possess it innately. We revised the abstract, introduction, and related work to clarify this distinction.

2. Relation to developmental visual diets

We now cite and discuss Lu et al., alongside Cai et al. and Vogelsang et al., in “Developmentally Motivated Visual Learning.” These studies motivate sensory restrictions based on developing acuity, contrast, and color sensitivity. Our study examines the consequences of early access to coherent object structure.

We also compare adaptations of their sensory transformations within our supervised STL10 setup. The DVD condition reaches 68.39% recognition accuracy and 52.73% shape bias, compared with 73.58% and 61.45% for Line→Photo. Blurred grayscale produces the highest shape bias, showing that this effect is shared with sensory restrictions. The full comparison appears in “Developmental sensory curricula.”

3. Training procedure and compute matching

The Method now states that Line→Photo first trains on drawings, then retains the learned model parameters and switches directly to the corresponding photographs. Supervised stages use the same class labels. For SSL, the optimizer and scheduler are reset at the switch; see “Self-Supervised Training and Evaluation.”

Photo-only matches the complete curriculum’s cumulative training epochs, including the initial line-drawing stage. In the original supervised recognition comparisons, shorter runs receive additional training after stagewise early stopping to equalize this budget; see “Supervised Learning Experiments.” The SSL experiments report matched 400-epoch comparisons and separate, explicitly reduced-budget runs.

4. Purpose of knowledge transfer

Distillation tests whether teachers with similar recognition accuracy differ in how effectively they transfer knowledge to smaller networks. In this experiment, the Line→Photo teacher reaches 69.01% accuracy using 70% of the data, while the Photo-only teacher reaches 68.89% using the full dataset. Students distilled from Line→Photo achieve higher accuracy across all three architectures. For example, MobileNetV1 reaches 75.48% versus 73.83%.

This suggests that the curriculum changes the usefulness of the teacher’s supervision. The more concentrated, structure-sensitive representations offer a possible explanation, although distillation alone does not establish that mechanism. We clarified this experiment’s purpose and moved it to the appendix under “Knowledge Distillation Experiments.”

Thank you for the response. The paper is improved and I increased my overall score. However, you should state more clearly that your paper isn't a developmental theory. There is still a lot of developmental motivation in the introduction, which may give the wrong impression.

Thank you for your careful consideration of our response and for increasing your score. We appreciate this clarification and agree that the scope should be clearer. We have revised the introduction to distinguish the developmental motivation from our computational investigation of early structural learning in artificial vision models. We are not proposing a developmental theory. Line drawings serve as a computational surrogate for structural representations, rather than for the retinal images infants experience.

Reviewer_kvVJOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates how structure-first learning, in which models are first trained on line drawings before transitioning to conventional photographs, shapes the learned visual representations. Through diverse downstream evaluations and analyses of the dynamics of network internal representations, the authors suggest that this training curriculum promotes more shape-centered and concentrated representations, while improving performance across multiple downstream tasks.

Critical strengths
  1. The paper is well-written and well-organized, it's easy to follow.
  2. The paper provides extensive evaluation covering diverse downstream tasks, complemented by detailed analyses of the network internal representation and its dynamics, offering insights into how such training recipe affects the learned representations.
Critical weaknesses

Line drawings introduce a strong structural and semantic prior by selectively removing semantically irrelevant texture and other visual details while preserving meaningful object contours. This raises two concerns:

1), it is questionable whether infants inherently possess such a prior during early visual development.

2), It is unclear whether the observed improvements (Line->Photo over Photo-only) arise simply from introducing this prior or from the developmental ordering of Line→Photo training, as the authors argue. This concern is further supported by the fact that Canny→Photo, which uses a simple edge detector without the semantic extraction capabilities of the learned line-drawing generator, shows little improvement over Photo-only training. An important control that trains on the same mixture of line drawings and photographs in a randomly shuffled, i.i.d. manner is needed to clarify this.

Questions for authors
  1. As mentioned in Weakness 2, could the author add a control that trains on the mixture of line drawings and photographs in a randomly shuffled, i.i.d. manner to isolate the effect of developmental ordering.

  2. Does the effect of early line-drawing training persist with longer photographic training? For example, if the proportion of line-drawing training epochs is varied, or photographic training is continued beyond the current training duration, would the observed representational changes persist or gradually be overwritten?

  3. How does the proposed curriculum compare with existing developmentally motivated training methods in following papers? A controlled comparison would help to justify the effectiveness of proposed method.

  • Lu, Z., Thorat, S., Cichy, R.M. and Kietzmann, T.C., 2026. Adopting a human developmental visual diet yields robust and shape-based AI vision. Nature Machine Intelligence, pp.1-14.
  • Vogelsang, M., Vogelsang, L., Pipa, G., Diamond, S. and Sinha, P., 2025. Potential role of developmental experience in the emergence of the parvo-magno distinction. Communications Biology, 8(1), p.987.
2 comments

We thank the reviewer for the thoughtful comments and positive assessment of the paper’s presentation and evaluations. The new controls help distinguish the contribution of structural input from the role of its timing.

1. Structural prior and developmental interpretation

We agree that the learned drawing generator supplies a strong semantic and structural prior. We do not assume that infants possess this prior innately or see the world as line drawings. Our developmental motivation concerns how early experience might establish structural representations that guide later learning. Drawings provide a simplified computational surrogate for those representations.

The weaker Canny results are consistent with the value of coherent object structure. Our new mixing controls examine how access to this structural information interacts with its timing.

2. Randomly interleaved and delayed-mixing controls

We added Photo+Line, which randomly interleaves drawings and photographs from the start, and Photo→Mixed, which introduces the mixture after initial photographic training. These comparisons use matched total training budgets.

Photo+Line achieves the highest recognition accuracy (75.19%), while Line→Photo produces the strongest shape bias (61.45%). Both outperform Photo→Mixed (69.09% recognition, 38.18% shape bias) and have higher mean neuronal structure sensitivity at all four stages. The overarching response provides the full comparison.

This result reveals an important distinction. An exclusive initial period of structural learning is not necessary for obtaining the benefits of structural experience. Rather, making structurally informative representations available from the beginning of training is sufficient to produce substantial improvements, and maintaining access to this information throughout training can further enhance recognition performance. We revised the manuscript accordingly; see “Early and delayed structural input.”

3. Persistence with longer photographic training

After 200 additional photographic epochs in both conditions, Line→Photo reaches 73.71% recognition accuracy and 46.61% shape bias, versus 69.13% and 14.50% for Photo-only. Shape bias decreases from 61.45% but retains a 32.11-point advantage. Thus, the structural preference weakens but persists over the tested continuation; see “Persistence during continued photographic training.”

4. Comparison with developmental sensory curricula

We now discuss Lu et al. and Vogelsang et al. in “Developmentally Motivated Visual Learning.” Within our supervised STL10 setup, we compare an adaptation of Lu et al.’s DVD filter, blurred grayscale inputs inspired by Vogelsang et al., and Cai et al.’s color-and-acuity transformations.

The DVD condition reaches 68.39% recognition accuracy and 52.73% shape bias. Blurred grayscale yields the highest shape bias, 67.27%, with 67.73% recognition accuracy. Line→Photo achieves the highest recognition accuracy among these conditions. These results show that structural preference can emerge through sensory restrictions as well as line-drawing experience. “Developmental sensory curricula” gives the full comparison.

Thank you for your response. The result of the interleaving baseline indicates that the improvement is not caused by the developmental order. Thus, it's a different story. And you should change it. I will keep my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Shared Concerns

We thank the reviewers for their thoughtful comments.

1. Developmental interpretation

We do not claim that line drawings realistically approximate infant vision or reproduce the developmental visual diet. Rather, our work is motivated by the hypothesis that early visual experience might induce structural representations in the visual cortex.

Sinha et al. (2026) propose that temporal coherence and motion-based grouping scaffold early perceptual organization. Temporal processing develops relatively rapidly compared with spatial acuity, potentially supporting object unity, figure-ground segregation, and boundaries. Infant and sight-restoration evidence supports this account.

From this perspective, line drawings provide a simplified computational surrogate for structural information that may be acquired through early perceptual organization, including object boundaries, global shape, and spatial configurations. We test whether early access to this structure establishes an inductive bias that persists during photographic learning. The pretrained generator supplies this prior; we do not assume it is innate in infants.

The benefits of Line→Photo persist across all three Gaussian-blur conditions on ImageNet-100, with gains of 0.60 to 1.27 percentage points over Photo-only. Thus, the benefit does not require precise, high-frequency contours, consistent with coarse structure supporting learning.

2. Related work and novelty

We now cite and discuss Vogelsang et al. (2018, 2025), Jinsi et al. (2023), Lu et al. (2026), and Cai et al. (2026).

Whereas developmental visual-diet studies investigate how sensory maturation shapes visual representations, our work examines how early learning from coherent object structure influences subsequent learning from richer visual input.

We additionally compare adaptations of the DVD filter, color-and-acuity transformations, and blurred grayscale inputs in our STL10 setup. Line→Photo achieves the highest recognition accuracy among these conditions, while blurred grayscale yields the highest shape bias.

Our principal contribution is to demonstrate the broader representational and functional consequences of early structural learning. Conventional cue-conflict shape-bias measures do not distinguish local from global shape sensitivity (Jarvers & Neumann, 2023). We therefore complement these measures with analyses of internal representations, depth prediction from frozen features, semantic segmentation, and knowledge distillation from teachers of approximately matched recognition accuracy.

3. Training budget and sample controls

Photo-only matches the complete Line→Photo curriculum's total training epochs, counting all stages.

Line→Photo adds line-drawing counterparts of the photographs, providing another representation of the same underlying object exemplars.

Photo→Line, Canny→Photo, blurred-photo→Photo, and the mixing controls examine this potential confound under matched training budgets. These controls show that input form and timing matter beyond simply adding another image representation.

4. Training order and persistence

Photo+Line randomly interleaves line drawings and photographs throughout training under a matched budget. Photo→Mixed introduces the mixture after initial photographic training.

CurriculumPhoto accuracy (%)Shape bias (%)
Photo-only68.7918.68
Line→Photo73.5861.45
Photo+Line75.1950.96
Photo→Mixed69.0938.18

This result reveals an important distinction. An exclusive initial period of structural learning is not necessary for obtaining the benefits of structural experience. Rather, making structurally informative representations available from the beginning of training is sufficient to produce substantial improvements, and maintaining access to this information throughout training can further enhance recognition performance. Photo→Mixed has lower recognition accuracy and shape bias than both curricula with early structural input, along with lower mean neuronal structure sensitivity at all four stages.

After 200 additional photographic epochs in both conditions, Line→Photo retains advantages of 4.58 percentage points in recognition and 32.11 points in shape bias over Photo-only. The structural preference weakens but persists.

These observations suggest two complementary roles for structural experience: structural initialization, in which early learning establishes a persistent inductive bias, and structural regularization, in which continued structural input may reinforce the representation during photographic learning.

Motion-derived structural cues and appearance information may likewise become available concurrently during development.

We have revised the manuscript to clarify that the essential finding is the benefit of early access to structurally informative visual experience, rather than the necessity or optimality of a strict structure-first training schedule.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.