Research map / Label-efficient and robust learning / Self-supervised and contrastive
Partial success in closing the gap between human and machine vision
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train a standard model on a much larger dataset | ImageNet-trained vision models · 17 OOD datasets | OOD accuracy: 0.559 -> 0.829 | Helped |
| Pre-train a vision transformer on 14M images | ViT-L · 17 OOD datasets | OOD accuracy: 0.706 -> 0.733 | Helped |
| Use self-supervised learning rather than standard supervision | ImageNet-trained CNNs · OOD datasets | OOD accuracy | No clear effect |
| Use SimCLR's training augmentations in supervised training (ablation) | SimCLR-matched supervised baselines · OOD datasets | OOD accuracy | Helped |
| Increase adversarial-training strength (ablation) | ResNet-50 · 17 OOD datasets | OOD accuracy: 0.545 -> 0.501 | Hurt |
| Increase adversarial-training strength (ablation) | ResNet-50 · Cue conflict images | shape bias | Helped |
| Train on large-scale datasets | Standard CNNs · 17 OOD datasets | error consistency | Helped |
| Train a ResNet-50 using CLIP hard or soft labels (ablation) | ResNet-50 · 17 OOD datasets | OOD accuracy | No clear effect |
| Use CLIP's joint image-text loss with a ResNet-50 backbone (ablation) | CLIP · OOD datasets | OOD accuracy | Hurt |
| Train the same ResNet-50 on Stylized-ImageNet rather than ImageNet (ablation) | ResNet-50 · Sketch images | error consistency: 0.74 -> 0.44 | Hurt |
| Replace ResNet-50 with ViT-S (ablation) | ImageNet-trained vision model · Sketch images | error consistency: 0.31 -> 0.67 | Helped |
Closest papers by meaning
- Human and AI Perceptual Differences in Image Classification Errors (AAAI 2025 · similarity 0.90)
- Evaluating Multiview Object Consistency in Humans and Image Models (NeurIPS 2024 · similarity 0.89)
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot (NeurIPS 2026 · similarity 0.87)
- Quantifying Uncertainty in Error Consistency: Towards Reliable Behavioral Comparison of Classifiers (NeurIPS 2025 · similarity 0.87)
- Contour Integration Underlies Human-Like Vision (ICML 2025 · similarity 0.87)
- Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision (NeurIPS 2026 · similarity 0.87)
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.