This paper proposes Lobster, an integrated framework for neurosymbolic learning that combines deep learning and symbolic reasoning. Lobster maps a common Datalog-based neurosymbolic language to the GPU programming paradigm, enabling end-to-end neurosymbolic learning on GPUs. It compiles to an intermediate language called APM, which supports discrete, probabilistic, and differentiable inference modes and implements optimization passes. Across nine applications across diverse domains, Lobster demonstrates an average speedup of 3.9x over the existing neurosymbolic framework, Scallop, enabling the scalability of previously infeasible tasks.
