Research map / Vision-language and multimodal / Human and embodied modeling
Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models
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.
Introduces the Visual Concept Connectome (VCC), an unsupervised open-world method that discovers concepts at every layer of a pretrained vision network and quantifies their interlayer connections via ITCAV.
Problem
Interpreting deep vision models by revealing human-interpretable concepts across layers and how concepts at one layer contribute to concepts at other layers, which prior single-layer concept methods cannot capture.
Approach
Builds a DAG of layer-wise concept clusters by recursively segmenting features conditioned on the layer above, then weights edges between concepts at different layers with Interlayer TCAV (ITCAV), a generalization of TCAV that needs no labelled concept images.
Key results
- Concept suppression over 50 ImageNet classes shows perturbing along discovered concepts reduces accuracy more than random directions for all tested models.
- Average path strength (ITCAV edge weights) correlates positively with logit sum scores for VGG16 and ViT-b, indicating predictive interlayer weights.
- Four-layer VCCs take about 15 minutes and all-layer VCCs about 36 hours on one NVIDIA Quadro RTX 6000 GPU; CNNs show decreasing branching factor while transformers keep it stable until the final layer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use model feature-space segmentation instead of layer-independent color segmentation | VCC segment proposals · ResNet50, VGG16, and ViT-b | segment size relative to receptive field | Helped |
| Suppress discovered concepts rather than a randomly selected direction | ImageNet classifiers · 50 randomly selected ImageNet classes | target-class accuracy under concept suppression | Hurt |
| Use ITCAV edge weights to connect concepts across layers | VCC · VGG16 and ViT-b | correlation between average path strength and logit sum | Helped |
| Use adversarially robust training on Stylized ImageNet | ResNet50 and VGG16 · ImageNet class VCCs | low-level concept count and compositionality | Hurt |
| Use SimCLR self-supervised training | ResNet50 and VGG16 · ImageNet class VCCs | low-level concept count and compositionality | Hurt |
| Use a transformer architecture rather than a CNN | ImageNet classifiers · Four-layer VCC analysis | final-layer edge-weight variance | Helped |
| Use a transformer architecture rather than a CNN | ImageNet classifiers · ImageNet classification | classification performance | Helped |
| Use concept-directed suppression rather than random perturbation | ImageNet classifiers · ImageNet class prediction | accuracy | Hurt |
| Represent concepts using more diverse connection strengths in the final layer | Transformer classifiers · ImageNet class VCCs | final-layer edge-weight variance | Helped |
Surprising findings
- Robust (adversarial) and self-supervised (SimCLR) models have fewer low-level concepts and less compositionality than the originals.
Limitations
- Validation is largely via proxy metrics (suppression, APS vs LS correlation) rather than human studies.
- Demonstrated only for image classification; extension to other tasks is stated but not tested.
Experimental setup
Datasets: ImageNet Broden
Benchmarks: ImageNet
Metrics: concept suppression accuracy average path strength vs logit sum correlation branching factor number of concepts edge weight average and variance receptive field vs relative concept size
Baselines: ACE
Builds on: TCAV ACE CRAFT maskSLIC
Closest papers by meaning
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.88)
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.87)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.87)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.86)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · similarity 0.85)
- A Convolutional Neural Network Interpretable Framework for Human Ventral Visual Pathway Representation (AAAI 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.