Neuro-symbolic AI improves neural networks by integrating symbolic structures like logical constraints, automata, and knowledge graphs. While neural components run efficiently on modern accelerators, symbolic components rely on sparse, irregular computations. These map poorly onto dense hardware like GPUs, creating performance bottlenecks despite their inherent parallelism.
This parallelism can be exposed by expressing symbolic structures as sparse tensor algebra. A compiler then divides these computations into independent units for hardware task schedulers, which dynamically exploit parallelism at run time. Combining this general software abstraction with flexible task-based hardware enables efficient execution of the symbolic half of neuro-symbolic AI, bringing the same hardware efficiency to irregular workloads that dense ones currently enjoy.
11/12/2026 11:00 - 12:00
ESAT Aula R
