Research map / Label-efficient and robust learning / Recognition and long-tail
Interpret Your Decision: Logical Reasoning Regularization for Generalization in Visual Classification
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 L-Reg, a sample-based logical (entropy-based) regularizer for visual classification that improves generalization in multi-domain generalization, generalized category discovery, and their combination.
Problem
Vision classifiers generalize poorly to unseen domains and to unknown classes; standard L2 weight decay is parametric rather than sample-based and gives ambiguous interpretability.
Approach
Derives a logic-based regularizer (L-Reg) from a logical analysis framework and semantic supports: an entropy-based loss that encourages each class to be decided by a minimal set of latent semantics with distinct supports per class, added to existing losses as L_main + alpha * L_L-Reg.
Key results
- On DomainBed with GMDG backbone, GMDG+L-Reg reaches 75.8 average accuracy versus 75.1 for GMDG (TerraIncognita 62.9 vs 60.7).
- On GCD with PIM, PIM+L-Reg improves unknown-class accuracy from 59.9 to 62.7 and average All accuracy from 67.4 to 68.8; adding Ortho-Reg gives 69.3 All.
- In the mDG+GCD setting with GMDG, L-Reg raises All accuracy from 47.94 to 51.94 and Unknown accuracy from 20.68 to 27.68.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add L-Reg to GMDG | GMDG · mDG average across PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet | accuracy: 75.1 -> 75.8; 0.7↑ | Helped |
| Add L-Reg to PIM | PIM · GCD average across six datasets | All accuracy: 67.4 -> 68.8; 1.4↑ | Helped |
| Add L-Reg to PIM | PIM · GCD average across six datasets | Known accuracy: 79.3 -> 79.0; 0.3↓ | Hurt |
| Add L-Reg to PIM | PIM · GCD average across six datasets | Unknown accuracy: 59.9 -> 62.7; 2.8↑ | Helped |
| Add L-Reg to GMDG | GMDG · mDG + GCD average | Unknown accuracy: 20.68 -> 27.68; 7.01 | Helped |
| Apply L-Reg to deep-layer features rather than earlier and deep layers (ablation) | GMDG · PACS mDG + GCD | All accuracy: 58.97 -> 67.82 | Helped |
| Add L-Reg to CircuitFormer | CircuitFormer · CircuitNet congestion prediction | pearson: 0.6374 -> 0.6553 | Helped |
| Add L-Reg to ERM | ERM · TerraIncognita mDG | accuracy: 49.9±3.6 -> 52.9±4.2 | Helped |
| Replace L-Reg with sparsity regularization (ablation) | PIM · GCD average across six datasets | All accuracy: 67.4 -> 66.6; -0.7 | Hurt |
| Replace L-Reg with Ortho-Reg (ablation) | PIM · GCD average across six datasets | All accuracy: 67.4 -> 68.4; 1.0 | Helped |
| Add Ortho-Reg alongside L-Reg (ablation) | PIM + L-Reg · GCD average across six datasets | All accuracy: 68.8 -> 69.3 | Helped |
| Add L-Reg to GMDG | GMDG · mDG benchmarks | accuracy: PACS 97.3±0.1 -> 97.4±0.2; VLCS 82.4±0.6 -> 82.4±0.0; OfficeHome 80.8±0.6 -> 80.9±0.5; TerraIncognita 60.7±1.8 -> 62.9±0 | Mixed |
Surprising findings
- L-Reg slightly lowers known-class accuracy in GCD (PIM known 79.3 to 79.0), a trade-off the authors attribute to reduced semantic support.
- Applying L-Reg to earlier layers improves unknown-class accuracy but degrades known-class accuracy, so it is recommended only for deep-layer features.
Limitations
- Known-class accuracy can slightly decrease because semantic supports are narrowed.
- Fails in strong domain shifts such as sketch images where person features are absent.
- Relies on the precondition that latent dimensions represent independent semantics, which may not hold for shallow layers.
Experimental setup
Datasets: PACS VLCS OfficeHome TerraIncognita DomainNet CIFAR10 CIFAR100 ImageNet-100 CUB Stanford Cars Herbarium19 CircuitNet
Benchmarks: PACS VLCS OfficeHome TerraIncognita DomainNet CIFAR10 CIFAR100 ImageNet-100 CUB Stanford Cars Herbarium19 CircuitNet
Metrics: accuracy All/Known/Unknown accuracy pearson spearman kendall
Baselines: ERM MIRO GMDG PIM GCD ORCA K-means CircuitFormer Ortho-Reg Sparsity
Builds on: GMDG PIM MIRO DomainBed
Closest papers by meaning
- DiVE-k: DIFFERENTIAL VISUAL REASONING FOR FINE-GRAINED IMAGE RECOGNITION (ICLR 2026 · similarity 0.86)
- Reasoning-Driven Multimodal LLM for Domain Generalization (ICLR 2026 · similarity 0.85)
- Democratizing Fine-grained Visual Recognition with Large Language Models (ICLR 2024 · similarity 0.84)
- Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation (ICLR 2026 · similarity 0.84)
- Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.83)
- Explaining Domain Shifts in Language: Concept Erasing for Interpretable Image Classification (CVPR 2025 · similarity 0.83)
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.