Research map / Label-efficient and robust learning / Recognition and long-tail
Concept frustration: Aligning human concepts and machine representations
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 concept frustration, a task-aligned Fisher-geometry diagnostic that detects when an unobserved concept induces contradictions among known concepts in a concept bottleneck model, validated on synthetic data and sarcasm and gull-vs-tern tasks.
Problem
Concept bottleneck models with incomplete ontologies lose accuracy and interpretability, and existing Euclidean comparisons between supervised concepts and SAE features fail to detect when a missing concept is needed.
Approach
Defines concept frustration as a sign contradiction among a triplet of two known concept directions and one unsupervised SAE direction, measured under a task-aligned Fisher information metric of a one-hidden-layer black-box classifier, with a closed-form Bayes-optimal CBM accuracy under a linear-Gaussian model.
Key results
- In the treasure hunter task, task-aligned frustration is higher on the round Earth than the flat Earth (paired Wilcoxon p=5.94e-3), while Euclidean frustration does not distinguish them (p=0.118).
- On synthetic data, CBM accuracy under frustrated regime (alpha=+1) is 0.699 vs 0.718 under consistent regime (alpha=-1), p<1e-4, while black-box accuracy is unchanged (p=0.132).
- Across sarcasm and CUB tasks, Fisher frustration of the (C1,C2) pair is higher for CBM1 than CBM2 (p=0.0488), whereas Euclidean frustration shows no difference.
Surprising findings
- Euclidean frustration does not detect frustration at all, even when the task-aligned metric does.
- Black-box accuracy does not drop under frustration, so frustration limits the CBM specifically rather than task difficulty.
Limitations
- Theory relies on a linear-Gaussian generative model, which may not hold for real representations.
- Pairwise frustration depends on choosing a maximally frustrating SAE direction, a heuristic choice.
- Task-aligned geometry is derived from a simple one-hidden-layer classifier, which may not generalise to richer models.
Experimental setup
Datasets: News Headlines sarcasm dataset CUB-200-2011 synthetic linear-Gaussian data
Benchmarks: News Headlines sarcasm CUB gull-tern synthetic treasure hunter
Metrics: accuracy global frustration pairwise frustration concept MSE semantic fidelity beta
Baselines: Euclidean similarity frustration black-box MLP sparse autoencoder
Builds on: concept bottleneck models sparse autoencoders
Closest papers by meaning
- Towards a learning theory of representation alignment (ICLR 2025 · similarity 0.86)
- Nonparametric Identification of Latent Concepts (ICML 2025 · similarity 0.85)
- Understanding Inter-Concept Relationships in Concept-Based Models (ICML 2024 · similarity 0.85)
- From Causal to Concept-Based Representation Learning (NeurIPS 2024 · similarity 0.85)
- Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.85)
- Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers (NeurIPS 2025 · 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.