Research map / Architectures and efficiency / CNN design
Alignment of CNN and Human Judgments of Geometric and Topological Concepts
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.
Evaluates whether ImageNet-pretrained CNNs show human-like sensitivity to 43 geometry and topology concepts from the Dehaene et al. odd-one-out task, comparing layers, depth, and class-level profiles to human groups.
Problem
Whether geometric and topological concepts that humans find intuitive are innate (core knowledge) or learned through visual experience; the paper tests if CNNs trained on natural images show human-like GT sensitivity.
Approach
Use ImageNet-pretrained CNNs without fine-tuning as proxies for visual experience, run them on the human odd-one-out GT task, and score odd-one-out via cosine-similarity ranking, then compare class-level accuracy profiles to human groups.
Key results
- ResNet-18 achieved the highest overall odd-one-out accuracy of 40%, versus 90.7% for Mundurucu adults and children and 62.8% for Western children ages 3-6.
- Class-level profile of ResNet-18 correlated r=.52 with Western children 3-6, r=.60 with Western adults, and r=.58 with Mundurucu; DenseNet correlated r=.81 with children 3-6.
- Increasing ResNet depth (18 to 152) did not improve alignment; ResNet-18 generally outperformed deeper variants, and all models showed a blind spot for Symmetrical Figures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ResNet-18 rather than the other tested CNNs | ImageNet-pretrained CNNs · Odd-one-out task, all 43 concepts | accuracy: 37.78% -> 40% | Helped |
| Use ResNet-18 rather than ResNet-34 | ResNet family · Odd-one-out task, all 43 concepts | accuracy: 24.44 -> 40 | Helped |
| Increase ResNet depth from 18 to 50 layers | ResNet family · Odd-one-out task, all 43 concepts | accuracy: 40 -> 33.33 | Hurt |
| Increase ResNet depth from 18 to 101 layers | ResNet family · Odd-one-out task, all 43 concepts | accuracy: 40 -> 33.33 | Hurt |
| Increase ResNet depth from 18 to 152 layers | ResNet family · Odd-one-out task, all 43 concepts | accuracy: 40 -> 37.78 | Hurt |
| Increase ResNet depth from 18 to 152 layers | ResNet family · Odd-one-out task, Geometrical Figures | accuracy: 33.33% vs. 55.56% | Helped |
| Increase ResNet depth from 18 to 152 layers | ResNet family · Odd-one-out task, Geometrical Transformations | accuracy: 12.5% vs. 25% | Helped |
| Use later rather than earlier ResNet-18 layers (ablation) | ResNet-18 · Odd-one-out task, most GT classes | average rank | Helped |
| Use later rather than earlier ResNet-18 layers (ablation) | ResNet-18 · Odd-one-out task, Metric Properties | average rank | No clear effect |
| Use later rather than earlier ResNet-18 layers (ablation) | ResNet-18 · Odd-one-out task, Geometrical Transformations | average rank | Hurt |
| Increase CNN depth within the ResNet family | ResNet models · Odd-one-out task, overall accuracy | accuracy: 40%; 24.44%; 33.33%; 33.33%; 37.78% | Mixed |
| Use ImageNet-pretrained CNNs on the odd-one-out task | CNNs · Odd-one-out task, all 43 concepts | accuracy: greater than 16.67%; ResNet-18 ... 40% | Helped |
Surprising findings
- Shallowest ResNet-18 outperformed deeper ResNet variants in overall accuracy and human alignment.
- Models performed below humans even though they were trained on large-scale image data, with all models near-blind to Symmetrical Figures concepts.
Limitations
- Models were not fine-tuned, so test stimuli were out-of-distribution, which partly explains poor performance.
- Small number of items (6 images per stimulus, 43 concepts) yields unstable per-class accuracies, with many classes having few items.
Experimental setup
Datasets: Dehaene et al. 2006 odd-one-out stimuli Izard and Spelke 2009 human data ImageNet (pretraining of models, not used by authors for training)
Benchmarks: Dehaene et al. 2006 odd-one-out task
Metrics: accuracy average rank of odd-one-out Pearson correlation point-biserial correlation
Baselines: Human adults and children (Mundurucu, Western) Campbell et al. 2024 CNN results
Builds on: Dehaene et al. 2006 odd-one-out task Izard and Spelke 2009 Marupudi and Varma 2023 Campbell et al. 2024
Closest papers by meaning
- Evaluating alignment between humans and neural network representations in image-based learning tasks (NeurIPS 2024 · similarity 0.87)
- Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests (NeurIPS 2026 · similarity 0.86)
- Feature Learning beyond the Lazy-Rich Dichotomy: Insights from Representational Geometry (ICML 2025 · similarity 0.86)
- Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning (CVPR 2025 · similarity 0.86)
- Position: The Platonic Representation Hypothesis (ICML 2024 · similarity 0.86)
- Measuring Vision-Language STEM Skills of Neural Models (ICLR 2024 · similarity 0.85)
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.