Research map / Label-efficient and robust learning / Recognition and long-tail
Formal Concept Lattices are Good Semantic Scaffolds for Concept-Based Learning
Deepika Vemuri, Sayanta Adhikari, Ankit Saha, Krishn Vishwas Kher, Vineeth N. Balasubramanian
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add iterative class-group refinement (ablation) | FoCA CBM · ImageNet100 | accuracy: 88.19 -> 91.92 | Improved |
| add iterative class-group refinement (ablation) | FoCA CBM · CIFAR100 | accuracy: 77.19 -> 79.7 | Improved |
| add iterative class-group refinement (ablation) | FoCA CBM · AwA2 | cluster impurity: 0.5231 -> 0.5288 | No clear effect |
| use joint instead of independent training (ablation) | FoCA CBM · ImageNet100 | accuracy: 77.54 -> 91.88 | Improved |
| use joint instead of sequential training (ablation) | FoCA CBM · ImageNet100 | accuracy: 86.69 -> 91.88 | Improved |
| supervise with random class groups instead of lattice groups (ablation) | FoCA CBM · ImageNet100 | accuracy: 91.36 -> 28.96 | Hurt |
| replace FCA lattice with GPT4-built hierarchy (ablation) | FoCA CBM · ImageNet100/AwA2/CIFAR100 | accuracy: 91.88 -> 91.47 (ImageNet100) | No clear effect |
| use FCA hierarchy instead of GPT4 hierarchy under 1% attribute removal (ablation) | FoCA CBM · ImageNet100 | accuracy drop (concept leakage): FCA 91.88 -> 60.57; GPT4 91.47 -> 89.20 | Improved |
| use hierarchical FoCA with lattice scaffolding on ViT | FoCA ViT · CIFAR100 | accuracy: 84.49 -> 86.65 | Improved |
| increase attribute loss weight alpha from 0.01 to 1 (ablation) | FoCA CBM · CIFAR100 | accuracy: 79.36 -> 73.50 | Hurt |
| increase group loss weight beta from 0.01 to 1 (ablation) | FoCA CBM · CIFAR100 | accuracy: 79.36 -> 65.09 | Hurt |
| stack all lattice attribute sets after backbone (naive) instead of staged supervision (ablation) | FoCA CBM-N · ImageNet100 | cluster impurity: 0.573 -> 0.665 | Hurt |
Closest papers by meaning
- From Tokens to Lattices: Emergent Lattice Structures in Language Models (ICLR 2025 · similarity 0.90)
- Learning Discrete Concepts in Latent Hierarchical Models (NeurIPS 2024 · similarity 0.86)
- Hyperbolic Concept Bottleneck Models (NeurIPS 2026 · similarity 0.86)
- Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models (CVPR 2024 · similarity 0.86)
- ICLR: In-Context Learning of Representations (ICLR 2025 · similarity 0.86)
- From Causal to Concept-Based Representation Learning (NeurIPS 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.
