Research map / Vision-language and multimodal / Human and embodied modeling
Learning from an Infant’s Visual Experience
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A doctoral consortium summary proposing to study how infant-like, instance-limited but viewpoint-rich visual experience (Toybox dataset) can support visual representation learning across three research questions.
Problem
Internet-style vision datasets are instance-rich and viewpoint-limited, lacking the repeated, multi-view regularities of infants' object-play visual experience, so models that exploit these regularities are scarce.
Approach
Characterize infant visual experience as instance-limited and viewpoint-rich and use egocentric Toybox videos to study self-supervised learning from multi-view object exposure, joint learning under distribution shift, and analogical reasoning over instance-viewpoint combinations.
Key results
- Viewpoint-equating self-supervised signal (modified SimCLR on Toybox) significantly outperforms baselines on downstream image classification of unseen Toybox objects, per prior CogSci 2023 work cited.
- Ongoing work finds categorical signals drive classification of unseen images while perceptual signals drive feature alignment across distributions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the modified embodied learning signal that equates different views of the same object | SimCLR · Toybox unseen objects | image classification performance | Helped |
| use categorical signals for joint learning from both distributions | joint learning under distribution shift · unseen images | classification performance | Helped |
| use perceptual signals for joint learning from both distributions | joint learning under distribution shift · two distributions | feature alignment | Helped |
Surprising findings
- Equating different viewpoints of an object causes the model to lose information about individual object geometry.
Limitations
- Doctoral consortium summary; experimental details and results are largely deferred to prior or ongoing work.
- Toybox-V2 annotations are not yet released.
Experimental setup
Datasets: Toybox
Builds on: SimCLR Toybox dataset
Closest papers by meaning
- Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning (CVPR 2025 · similarity 0.88)
- BabyVLM-V2: Toward Developmentally Grounded Pretraining and Benchmarking of Vision Foundation Models (CVPR 2026 · similarity 0.85)
- BabyVision: Visual Reasoning Beyond Language (ICML 2026 · similarity 0.84)
- Alignment of CNN and Human Judgments of Geometric and Topological Concepts (AAAI 2025 · similarity 0.83)
- Object Concepts Emerge from Motion (NeurIPS 2025 · similarity 0.83)
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.