Projects

A Standard Codebase for Interpretable Deep Learning

An open-source library that gives interpretable deep learning a common foundation. By standardizing how interpretable models are implemented, it aims to accelerate development, make benchmarks fairer and more comprehensive, and open research opportunities beyond what current approaches allow.

GitHub repo      Documentation

A Theory of Interpretable Deep Learning

Interpretability still lacks a general formal theory, which has left the field fragmented and its evaluation protocols inconsistent. This project develops a principled framework that begins from explicit premises about what interpretability means for a target user and systematically derives the constraints an interpretable model must satisfy. The aim is to unify traditional, concept-based, and mechanistic interpretability under a single theory.

Structured Interpretable Models

Adding explicit structure to the internal representations of interpretable models. Instead of treating interpretable variables such as concepts as an unstructured set, this project organizes them through causal graphs, probabilistic graphical models, and ontologies, so that models reason in ways that are more faithful, modular, and aligned with domain knowledge.

Interpretability for Deep Time Series Models

Extending interpretability-by-design to deep models that reason over time. This project investigates what interpretability should mean for temporal data and develops models whose predictions are expressed through meaningful variables and mechanisms, and that remain interpretable as the underlying processes evolve over time.