Advances in Interpretable Machine Learning

The field of machine learning is moving towards more interpretable and explainable models. Recent developments have focused on creating frameworks that can provide insights into the decision-making processes of complex models. One of the key directions is the use of kernel methods, which provide a theoretically grounded framework for non-linear and non-parametric learning. Another area of research is the development of methods for analyzing and interpreting the features used by trained neural networks. Noteworthy papers in this area include the introduction of the loss kernel, a geometric probe for deep learning interpretability, and the use of the empirical neural tangent kernel to surface the features used by trained neural networks. Additionally, there have been significant advances in the development of methods for imputing missing data, including the use of tensor trains and implicit neural representations.

Sources

Kernel Regression of Multi-Way Data via Tensor Trains with Hadamard Overparametrization: The Dynamic Graph Flow Case

Interpretable Kernel Representation Learning at Scale: A Unified Framework Utilizing Nystr\"om Approximation

Meta-Learning Theory-Informed Inductive Biases using Deep Kernel Gaussian Processes

The Loss Kernel: A Geometric Probe for Deep Learning Interpretability

Bayesian Influence Functions for Hessian-Free Data Attribution

Enhancing Certifiable Semantic Robustness via Robust Pruning of Deep Neural Networks

Feature Identification via the Empirical NTK

TabINR: An Implicit Neural Representation Framework for Tabular Data Imputation

Diffusion Transformers for Imputation: Statistical Efficiency and Uncertainty Quantification

Built with on top of