Tractable Neuro-Causal Models

Modern artificial intelligence (AI), especially deep learning, has achieved remarkable success but it also has serious drawbacks. These systems often need vast amounts of data, are difficult to interpret, and can easily make mistakes in unfamiliar situations or when facing unexpected inputs. That makes them risky to use in critical areas like healthcare, finance, or justice. One promising way to address these weaknesses is by combining deep learning with causal reasoningan approach that mirrors how humans think about cause and effect. Causal models can help AI understand why something happens, not just what is likely. This opens the door to systems that generalize better, adapt to new situations, and make more trustworthy decisions. However, a significant challenge in combining deep learning with causality is inference. In deep learning, getting an answer is fast and direct, while causal reasoning requires probabilistic inferencebasically, the process of computing huge sums over differently weighted possibilities which is in general a computationally very hard task. In this project, we aim to develop a new type of model, called causal circuits, built on probabilistic circuitsa special kind of neural-like architecture designed for tractable probabilistic inference. In simple terms, tractability means these models can perform complex reasoning tasks both exactly and efficiently, even when dealing with high-dimensional data. This is a game-changer: it makes it possible to answer what if and why questions in real time, without needing shortcuts or approximations. The goals of the project are: Develop causal circuits that allow exact, tractable reasoning about cause and effect; Create methods to learn these circuits directly from data, even when the underlying causal structure is unknown; Combine causal circuits with powerful deep learning tools to form hybrid AI systems that are both expressive and computationally efficient. Ultimately, this research could pave the way for a new generation of AIsystems that are not just smart, but also reliable, understandable, and robust.