DeepSWIP: Quotient-WMC Counterfactuals for Neural Probabilistic Logic Programs

DeepSWIP is a novel framework that introduces a single-world counterfactual semantics for DeepProbLog programs, enhancing standard associational inference with causal reasoning capabilities. It leverages neural materialization to convert fixed-context neural predicates into conventional ProbLog choices, enabling the computation of counterfactuals through weighted model counting (WMC) under specific assumptions. Experimental results demonstrate a 2.14× speedup in inference tasks on MPI3D, highlighting its potential for improving efficiency and accuracy in neural probabilistic logic applications, particularly in scenarios requiring causal interventions.