This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper presents a novel approach leveraging GPU acceleration to overcome the limitations of conventional CPU-based logic programming, which leads to poor performance in large-scale inference tasks. Leveraging the high-throughput matrix computation capabilities of GPUs, we propose two GPU-accelerated Boolean Matrix Logic Programming (BMLP) algorithms for top-down inference on linear binomial cyclic datalog programs. We extend the BMLP theoretical framework to support general linear recursion using binary predicates. Experimental results on reachability queries for large-scale directed graphs and the Freebase 15K dataset demonstrate that they achieve one to four orders of magnitude speedup over existing state-of-the-art systems. This demonstrates that Boolean matrix-based inference can significantly enhance the scalability and efficiency of logic programming on modern hardware. The source code is available at https://github.com/lun-ai/BMLP.git .
We demonstrate that the BMLP algorithm utilizing GPU acceleration can dramatically improve the performance of large-scale logic programming.
◦
Contributes to solving the scalability problem of logic programming by achieving 1 to 4 orders of magnitude speed improvement over existing systems.
◦
We experimentally verify the efficiency of Boolean matrix-based inference and present a performance evaluation using a real dataset.
◦
Ensure reproducibility and facilitate utilization by other researchers through open source source code.
•
Limitations:
◦
Currently, the algorithm is limited to linear binomial recursion and binary predicates, and scalability to more complex logic programs is required.
◦
Since this is an optimization for a specific type of datalog program, further research is needed to determine whether it can be applied to all types of logic programming problems.
◦
Since the experiment is limited to a specific dataset, generalizability to various datasets must be verified.
