Towards Deep-Learning Architectures That Respect Causal Principles

Deep learning and causal inference have largely developed as separate fields. Deep learning provides scalability and flexibility, but it mainly learns statistical associations between variables. Causal learning, on the other hand, aims to understand the underlying mechanisms that generate the data, but it can be computationally expensive and difficult to scale. When the two fields are used sequentially e.g. when one is only used to guide or validate the other, both can lose some of their main strengths. The key question is therefore: Can causal structure serve as a predictive principle for neural architecture design, rather than merely as a constraint imposed during or after learning?