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

Explain Yourself, Briefly! Self-Explaining Neural Networks with Concise Sufficient Reasons

ICLR 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Post-hoc explanation analysis · 0 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 sufficient subset training (SST), a self-supervised training method that makes neural networks output concise sufficient reasons alongside predictions, with baseline, probabilistic, and robust masking, and proves complexity/approximation hardness of post-hoc minimal sufficient reasons.

Problem

Post-hoc computation of cardinally minimal sufficient reasons for neural network predictions is computationally intractable and relies on out-of-distribution input assignments, yielding suboptimal or counterintuitive subsets.

Approach

Train the model with an extra sigmoid output that selects a feature subset, using dual propagation with a faithfulness loss (prediction consistency under masked complement) and an L1 cardinality loss, so explanations are produced inherently and the model is exposed to mixed inputs during training.

Key results

• On MNIST, SST with robust masking yields explanations of 1.42% average size with 99.28% robust faithfulness, versus 2.35% and 98.08% for SIS, at about 1e-6 seconds per explanation.
• On IMDB, B-SST achieves 98.05% baseline faithfulness versus 23.37% for S-Anchors, with explanation time around 1e-3 seconds versus roughly 20 seconds for Anchors.
• Accuracy changes are small: MNIST +0.15%, CIFAR-10 -1.96%, ImageNet -4.7%, IMDB -0.87%/-0.89%.

Measured effects reported in the paper

Surprising findings

• Baseline masking with a fixed zero baseline on MNIST led to larger subsets (23.69%) and a larger accuracy drop, attributed to bias toward OOD inputs.
• Language SST subsets are larger than post-hoc subsets (e.g., 28.96% vs 1.12% for Anchors on IMDB), trading size for much higher faithfulness.

Limitations

• Requires tuning additional hyperparameters (threshold, faithfulness and cardinality coefficients), with grid search only over the cardinality coefficient.
• Dual-propagation training is about twice as computationally demanding as standard training, more so for robust masking.
• Overly optimizing interpretability constraints can reduce accuracy, as seen on ImageNet.

Experimental setup

Datasets: MNIST CIFAR-10 ImageNet IMDB SNLI

Benchmarks: MNIST CIFAR-10 ImageNet IMDB SNLI

Metrics: explanation size time robust faithfulness probabilistic faithfulness baseline faithfulness timeouts accuracy gain

Baselines: Anchors SIS gradient search S-Anchors

Builds on: Anchors SIS PGD adversarial training BERT

Closest papers by meaning

• Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.89)
• Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.88)
• Towards Modeling Uncertainties of Self-Explaining Neural Networks via Conformal Prediction (AAAI 2024 · similarity 0.87)
• Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 2025 · similarity 0.86)
• How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.86)
• Q-SENN: Quantized Self-Explaining Neural Networks (AAAI 2024 · similarity 0.85)