Tractable Neuro-Causal Models

• Applicant: Professor Dr. Kristian Kersting
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 545991504

Wider research context. Deep learning, the main driving force behind artificial intelligence (AI), suffers from data-hungriness, lack of interpretability and explainability, as well as proneness to adversarial attacks and shortcut learning. A principled approach to overcome these weaknesses is causal modeling and inference, a mathematical framework well aligned with human-like cognition. Indeed, unifying causality and deep learning has been identified as one of the most promising avenues towards trustworthy and human-aligned AI. However, a genuine combination of causality and deep learning poses significant computational challenges, due to their fundamentally different ways of performing inference. Causal inference, in particular, requires probabilistic inference as a subroutine, which is a notoriously hard computational problem. Objectives. Our overarching goal is to develop models that combine deep learning with principled causal reasoning, in order to take a big step towards scalable neuro-causal AI-systems that are expressive, robust, and trustworthy. A particular focus in this project is on the computational challenge of inference. In contrast to existing neuro-causal models, which are usually based on approximate probabilistic inference, we will base our models on tractable probabilistic circuits (PCs), which perform probabilistic inference exactly and efficiently. The main objectives are (i) generalizing PCs to the causal realm, leading to so-called causal circuits (CCs), and (ii) combining CCs with probabilistic deep learning models via hybrid inference schemes. Approach. To establish CCs, we will first adopt structural causal models (SCMs) via compilation, i.e., we will convert SCMs into tractable circuits and study causal inference within these. In contrast to existing approaches, which are restricted to discrete data, our compilation methods will also handle continuous data. Furthermore, we will develop algorithms to learn CCs directly from data and study them as a “first-class” causal model, generalizing SCMs. Moreover, in order to connect CCs with deep learning models, we will build on recent hybrid probabilistic inference techniques to combine exact CC inference with approximate inference in probabilistic deep learning systems.

• DFG Programme: Research Grants
• International Connection: Austria
• Cooperation Partner: Professor Dr. Robert Peharz