Towards Scalable Multimodal Causal Deep Learning: An Exploratory Study

We’ve received funding from the Hasler Fooundation on the project: “Towards Scalable Multimodal Causal Deep Learning: An Exploratory Study”!

Abstract Contemporary deep learning (DL) models excel in approximating complex functions by discovering fine-grained correlations among features. Yet, DL models struggle to generalize in out-of-distribution settings as they fall short in identifying causal relationships. This small project explores the possibility to construct a unified DL architecture supporting causal inference queries and thus producing more robust, explainable, and generalisable predictions. This architecture will combine Retrieval Augmented Generative techniques to learn a causal graph from the analysis of relevant scientific literature, and Deep Concept Reasoning models to infer structural equations making the underlying causal mechanisms explicit. The project will perform a preliminary exploration of this approach, with the aim of both testing its feasibility and drafting an SNFS grant proposal to further develop and scale it to real-world applications.