Research map / Label-efficient and robust learning / Recognition and long-tail

Q-SENN: Quantized Self-Explaining Neural Networks

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Self-explaining and visual explanation networks · 12 citations (OpenAlex)

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 Q-SENN, a self-explaining network whose final layer is a sparse ternary (-1/0/+1 scaled) feature-class assignment, built via iterative quantized sparse fine-tuning, plus a CLIP-based method to align learned features with attribute concepts without annotations.

Problem

Self-Explaining Neural Networks (SENN) produce concept-based explanations but their earlier realizations lose accuracy and do not scale beyond simple datasets like MNIST; concepts also tend to be class detectors rather than shared human-understandable features.

Approach

Quantize the sparse final layer of a SLDD-style model to ternary values (0, ±alpha) and alternate computing this sparse quantized layer with fine-tuning the backbone for several iterations, so features become more general, binary, diverse and grounded while keeping about 5 features per class.

Key results

• On CUB-2011 with ResNet-50, Q-SENN reaches 85.9% accuracy (dense 86.7%) and Loc@5 of 78.1% vs 67.2% for SLDD-Model.
• On TravelingBirds (spurious background shift), Q-SENN reaches 67.3% accuracy vs 64.1% for SLDD-Model and 39.4% for dense ResNet.
• Dependence gamma on CUB drops to 30.0% vs 46.9% for SLDD-Model, with about 5 features per class and 50 total features.

Measured effects reported in the paper

Surprising findings

• Sparser, more interpretable models (fewer features per class) gave higher accuracy on TravelingBirds, suggesting the dense model learns spurious correlations.
• Quantization increased Loc@5 and binary-feature fraction but reduced accuracy on TravelingBirds compared with the non-quantized ablation (70.0%).

Limitations

• Human-subject experiments are needed to verify interpretability for humans; not performed.
• Alignment with human concepts depends on CLIP and only validated on CUB attributes.
• Iterative procedure costs more training time than the SLDD-Model baseline.

Experimental setup

Datasets: CUB-2011 FGVC-Aircraft Stanford Cars TravelingBirds ImageNet

Benchmarks: CUB-2011 FGVC-Aircraft Stanford Cars TravelingBirds ImageNet

Metrics: accuracy Loc@5 alignment r dependence gamma fraction of binary features

Baselines: Dense ResNet glm-saga_5 Concept Bottleneck Model (joint) SLDD-Model PIP-Net ProtoPool

Builds on: Self-Explaining Neural Networks SLDD-Model glm-saga CLIP Feature Diversity Loss

Closest papers by meaning

• Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.87)
• Towards Modeling Uncertainties of Self-Explaining Neural Networks via Conformal Prediction (AAAI 2024 · similarity 0.86)
• QPM: Discrete Optimization for Globally Interpretable Image Classification (ICLR 2025 · similarity 0.86)
• Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 2025 · similarity 0.85)
• Gnothi Seauton: Empowering Faithful Self-Interpretability in Black-Box Transformers (ICLR 2025 · similarity 0.85)
• Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.85)