Beyond Differentiability: Neurosymbolic Learning with Black-Box Programs
Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: learning algorithms, symbolic reasoning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Neurosymbolic learning has demonstrated promising potential as a paradigm to combine the worlds of classical algorithms and deep learning. However, existing general neurosymbolic frameworks require that programs be written in differentiable logic programming languages, restricting their applicability to a small fragment of algorithms. We introduce Infer-Sample-Estimate-Descend (ISED), a general algorithm for neurosymbolic learning with black-box programs. We evaluate ISED extensively on a set of 30 benchmark tasks that encompass rich data types and reasoning patterns. ISED achieves 30% higher accuracy than end-to-end neural baselines. Moreover, ISED's solutions often outperform those obtained using Scallop, a state-of-the-art neurosymbolic framework: the programs in 17 (61%) of the benchmarks cannot be specified using Scallop, and ISED on average achieves higher accuracy on those that can be specified using Scallop.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7948
Loading