Reproduce “REASON: Accelerating Probabilistic Logical Reasoning for Scalable Neuro-Symbolic Intelligence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zishen Wan, Che-Kai Liu, Jiayi Qian, Hanchen Yang, Arijit Raychowdhury, Tushar Krishna
REASON introduces a co-designed hardware-software framework to accelerate probabilistic logical reasoning within neuro-symbolic AI systems. This specialized architecture and set of algorithmic optimizations achieve 12-50x speedup and 310-681x energy efficiency compared to GPUs on representative neuro-symbolic workloads, enabling real-time operation at 2.12 W average power.
Key insights
- Probabilistic logical reasoning kernels, despite their diverse functions, can be abstracted into a unified Directed Acyclic Graph (DAG) representation, enabling consistent optimization and hardware mapping.
- The inherent irregularity, memory-boundedness, and low arithmetic intensity of symbolic and probabilistic workloads necessitate a dedicated hardware-software co-design for efficient execution.
- Optimizing for sparse, irregular memory accesses and fine-grained data dependencies through architectural features like reconfigurable tree engines and hardware-managed watched literals is critical for accelerating symbolic inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.