Research map / Label-efficient and robust learning / Recognition and long-tail
Human Cognition-Inspired Hierarchical Fuzzy Learning Machine
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.
Proposes HC-HFLM, a fuzzy learning machine that uses a hierarchical alignment loss to align data-derived fuzzy similarity hierarchies with class-knowledge hierarchies (from WordNet, expert or LLM class descriptions), with a quotient-space theory proof and gains on six classification benchmarks.
Problem
Standard classifiers assume crisp class concepts and ignore relations between classes from human knowledge, limiting interpretability and generalization.
Approach
Represent class knowledge as a fuzzy similarity relation on the class space and train a sample-level fuzzy similarity relation with a hierarchical alignment loss so the data's quotient-space hierarchy matches the knowledge hierarchy; the loss is shown theoretically to induce this alignment.
Key results
- CK1-HFLM reaches 95.06% on FLO and 68.78% on CUB, versus 94.05% and 66.19% for FLM.
- CK1-HFLM achieves 91.10% on AWA1 and 93.59% on AWA2, above FLM's 89.95% and 92.79%.
- CK2-HFLM (WordNet) reaches 76.20% on ImageNet1K, the best among compared methods, and 90.21% on APY.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Integrate class description vectors | FLM · APY test | accuracy: 89.42 -> 90.23 | Helped |
| Integrate WordNet knowledge | FLM · APY test | accuracy: 89.42 -> 90.21 | Helped |
| Integrate class description vectors | FLM · ImageNet1K test | accuracy: 76.02 -> 76.16 | Helped |
| Integrate WordNet knowledge | FLM · ImageNet1K test | accuracy: 76.02 -> 76.20 | Helped |
| Integrate class description vectors | FLM · AWA1 test | accuracy: 89.95 -> 91.10 | Helped |
| Integrate WordNet knowledge | FLM · AWA1 test | accuracy: 89.95 -> 90.87 | Helped |
| Integrate class description vectors | FLM · AWA2 test | accuracy: 92.79 -> 93.59 | Helped |
| Integrate WordNet knowledge | FLM · AWA2 test | accuracy: 92.79 -> 93.35 | Helped |
| Integrate class description vectors | FLM · FLO test | accuracy: 94.05 -> 95.06 | Helped |
| Integrate class description vectors | FLM · CUB test | accuracy: 66.19 -> 68.78 | Helped |
Surprising findings
- WordNet-based knowledge outperforms class description vectors on ImageNet1K, while domain-expert description vectors win on APY and AWA.
Limitations
- Gains on the six benchmarks are modest (typically about 1 point over FLM).
- Class knowledge quality must be judged manually and refined by case (high, medium, low), limiting automation.
- Experiments use pre-extracted features with fully connected backbones rather than end-to-end training.
Experimental setup
Datasets: MNIST APY ImageNet1K AWA1 AWA2 FLO CUB WordNet class description vectors
Benchmarks: APY ImageNet1K AWA1 AWA2 FLO CUB
Metrics: test accuracy
Baselines: KNN DT SVM NB CEC FLM
Builds on: Fuzzy learning machine (FLM) Knowledge theory of concepts (Murphy & Medin) Quotient space theory (Zhang & Zhang) Fuzziness permissible loss
Closest papers by meaning
- Human-like Category Learning by Injecting Ecological Priors from Large Language Models into Neural Networks (ICML 2024 · similarity 0.81)
- To Each Metric Its Decoding: Post-Hoc Optimal Decision Rules of Probabilistic Hierarchical Classifiers (ICML 2025 · similarity 0.80)
- MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · similarity 0.79)
- Learning Discrete Concepts in Latent Hierarchical Models (NeurIPS 2024 · similarity 0.79)
- Generating Computational Cognitive models using Large Language Models (NeurIPS 2025 · similarity 0.79)
- Can Large Language Models Derive High-Level Cognition from Low-Level and Fragmented Foundational Information? (AAAI 2025 · similarity 0.78)
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.