Interpretable Deep Learning: From Theory to Frontier Models
Our tutorial slides, materials, and a detailed schedule will be posted here as soon as we finalize them.
For a complete bibliography of the topics and works discussed in this tutorial, please refer to our resources section.
Motivation
As frontier models grow more complex and autonomous, interpretability has become essential for understanding, debugging, and controlling them. Interpretable Deep Learning (IDL) adds structural priors to a model’s architecture, data, or training objective. These priors make models auditable, aligned with human values, compliant with security protocols, and monitorable at scale.
Objectives
This tutorial prepares AI researchers and engineers to understand IDL principles and interfaces and to build on the state of the art. We first derive from first principles how to design IDL models and standardized IDL APIs. We then show how standardized data structures, layers, losses, and interventions let researchers reproduce, compare, and extend existing methods, using PyTorch Concepts to steer, align, and debug models. Next, we scale these ideas to language models: we build interpretable LMs from scratch and fine-tune frontier models for interpretability and safety with BabySteerling. We highlight open challenges throughout and close with a discussion of IDL for frontier-model monitoring and security, data annotation, and deployment in safety-critical settings.
Target Audience
The tutorial targets researchers in deep learning and in intrinsic and mechanistic interpretability, researchers from other ML subfields, and practitioners building applications where interpretability is paramount, such as healthcare. To serve these groups, we start with a shared background, then introduce IDL APIs, and then turn to safety use cases of interpretable LMs. Each technical part ends with a Q&A with the presenters.
Required Background
Our material will assume a basic knowledge of ML (e.g., foundations of supervised learning, experimental design, basic probabilistic modelling, etc.), with particular emphasis on a solid Deep Learning foundation (e.g., tensor calculus, neural networks, backpropagation, etc.). Concepts that may require mathematical tools/expertise beyond those one would expect to be shared among the AAAI community will be (re)introduced in our tutorial.
Lab Requirements
The technical parts include hands-on coding demonstrations with PyTorch Concepts and BabySteerling. In the style of the PyTorch Lightning tutorials, coding slides accompany the technical slides to show how each theoretical component translates into code. Our accompanying website already hosts our previous AAAI and UAI tutorials and a list of reference papers and repositories. Before the tutorial, we will add all slides and notebooks to the site. Afterward, if a recording is available, we will post it synced with a time-stamped slide deck.
Presenters
Citing This Tutorial
If you found this tutorial useful for your research, blogs, or work, please cite it as follows:
Abdelsalam Ismail A., Barbiero P., and De Felice G. (2027). Interpretable Deep Learning: From Theory to Frontier Models. Tutorial at the Forty-First Annual AAAI Conference on Artificial Intelligence (AAAI). https://interpretabledeeplearning.github.io/
Or use the following bibtex entry:
@misc{interpretabledl2027tutorial,
title = {Interpretable Deep Learning: From Theory to Frontier Models (AAAI-2027)},
author = {Abdelsalam Ismail, A. and Barbiero, P. and De Felice, G.},
year = {2027},
howpublished = {Tutorial at the Forty-First Annual AAAI Conference on Artificial Intelligence (AAAI)},
url = {https://interpretabledeeplearning.github.io/}
}