Interpretability

Zipper: A novel statistical device resolves the degeneracy issue in algorithm-agnostic inference, enabling reliable goodness-of-fit tests with enhanced power.

Researchers developed RUM, a refined unlearning meta-algorithm, that significantly improves existing unlearning methods by strategically refining forget sets and employing appropriate unlearning algor…

AI explanations can be subtly manipulated to influence human decisions, highlighting the urgent need for more robust and ethical AI explanation design.

Noise in data surprisingly simplifies machine learning models, improving their interpretability without sacrificing accuracy; this paper quantifies this effect across various hypothesis spaces.

R2ET: training for robust ranking explanations by an effective regularizer.

DNNs learn interactions in two phases: initially removing complex interactions, then gradually learning higher-order ones, leading to overfitting.

GNAN: a novel interpretable graph neural network achieving accuracy comparable to black-box models.

This paper presents nearly-linear time algorithms for testing model calibration, improving upon existing methods and providing theoretical lower bounds for various calibration measures.

Stochastic Concept Bottleneck Models (SCBMs) revolutionize interpretable ML by efficiently modeling concept dependencies, drastically improving intervention effectiveness and enabling CLIP-based conce…

Stochastic Amortization accelerates feature and data attribution by training amortized models using noisy, yet unbiased, labels, achieving order-of-magnitude speedups over existing methods.

RegExplainer unveils a novel method for interpreting graph neural networks in regression tasks, bridging the explanation gap by addressing distribution shifts and tackling continuously ordered decisio…

Optimal ablation (OA) improves model interpretability by precisely measuring component importance, outperforming existing methods. OA-based importance shines in circuit discovery, factual recall, and …

One-Sample-Fits-All (OFA) framework efficiently approximates all probabilistic values simultaneously, achieving faster convergence rates than existing methods.

DNNs are powerful but lack the clear semantics of PGMs. This paper innovatively constructs infinite tree-structured PGMs that exactly correspond to DNNs, revealing that DNN forward propagation approxi…

Multilinear Mixture of Experts (μMoE) achieves scalable expert specialization in deep neural networks through tensor factorization, enabling efficient fine-tuning and interpretable model editing.

Adaptive greedy algorithms significantly improve the accuracy of identifying the most influential subset of training data, overcoming limitations of existing methods that fail to capture complex inter…

Counterfactual Clamping Attack (CCA) improves model reconstruction using counterfactual explanations by leveraging decision boundary proximity, offering theoretical guarantees and enhanced fidelity.

New metrics and p-annealing improve sparse autoencoder training for better language model interpretability.

New scalable method measures per-unit interpretability in vision DNNs without human evaluation, revealing anti-correlation between model performance and interpretability.

MambaLRP enhances explainability of Mamba sequence models by ensuring faithful relevance propagation, achieving state-of-the-art explanation performance, and uncovering model biases.