Interpretability

Unlocking complex models’ secrets: New algorithm identifies input interactions using the Möbius Transform, boosting interpretability with surprising speed and accuracy.

This paper introduces a novel framework for learning discrete concepts from high-dimensional data, establishing theoretical conditions for identifying underlying hierarchical causal structures and pro…

Interpretable Mesomorphic Neural Networks (IMNs) achieve accuracy comparable to black-box models while offering free-lunch explainability for tabular data through instance-specific linear models gener…

M-GAM: Interpretable additive models handling missing data with superior accuracy & sparsity!

CMR: A novel Concept-Based Memory Reasoner delivers human-understandable, verifiable AI task predictions by using a neural selection mechanism over a set of human-understandable logic rules, achievin…

Boosting machine learning model interpretability, this paper introduces cluster-based and tree-based Sparse Explanation Values (SEV) for generating more meaningful and credible explanations by optimiz…

End-to-end sparse autoencoders revolutionize neural network interpretability by learning functionally important features, outperforming traditional methods in efficiency and accuracy.

GRAPHTRAIL unveils the first end-to-end global GNN explainer, translating black-box GNN predictions into easily interpretable boolean formulas over subgraph concepts, achieving significant improvement…

Edge Pruning efficiently discovers sparse, yet accurate, computational subgraphs (circuits) in large language models via gradient-based edge pruning, advancing mechanistic interpretability research.

New method, STRIPE-X, powerfully detects ’encoding’ in AI explanations—a sneaky phenomenon where explanations predict outcomes better than their constituent parts alone would suggest.

Chess AI Leela Zero surprisingly uses learned look-ahead, internally representing future optimal moves, significantly improving its strategic decision-making.

Dual-Perspective Activation (DPA) efficiently denoises ANN channels by jointly using forward and backward propagation criteria, improving sparsity and accuracy.

Researchers dissected attention paths in Transformers using statistical mechanics, revealing a task-relevant kernel combination mechanism boosting generalization performance.

Denoising Diffusion Path (DDPath) uses diffusion models to dramatically reduce noise in attribution methods for deep neural networks, leading to clearer explanations and improved quantitative results.

Data-faithful feature attribution tackles misinterpretations from unobservable confounders by using instrumental variables to train confounder-free models, leading to more robust and accurate feature …

Researchers developed a novel method using mechanistic interpretability to create compact formal proofs for AI model performance, improving AI safety and reliability.

Causal Dependence Plots (CDPs) visualize how machine learning model predictions causally depend on input features, overcoming limitations of existing methods that ignore causal relationships.

This paper presents a novel method to make black box neural networks intervenable using only a small validation set with concept labels, improving the effectiveness of concept-based interventions.

B-cosification: cheaply transform any pre-trained deep neural network into an inherently interpretable model.

Auditing local explanations is surprisingly hard: proving explanation trustworthiness requires far more data than previously thought, especially in high dimensions, challenging current AI explainabil…