Characterization of Neuro-Symbolic AI and Graph Convolutional Network workloads
Cory Davis, Patrick Stockton, Eugene B John, Zachary Susskind, Lizy K. John · 2024
The explosive growth of artificial intelligence has created new domains of AI models. These domains include Neuro-Symbolic AI (NSAI) and Graph Neural Networks (GNN). NSAI and GNN models have already demonstrated the capability to significantly outperform deep learning models in domains such as image and video reasoning, and network classification, respectively. They have also been shown to obtain high accuracy with significantly less training data than traditional neural network models. However, the recent emergence of the field, and relative sparsity of published results, leads to a meager understanding of the performance characteristics of these models. In this work, we describe and analyze four models in the NSAI and GNN domains. We find that the NSAI models have less potential for parallelism than traditional neural models due to complex control flow, low compute-to-byte operations, and high cost of data movement. Additionally, in the graph network, we find an abundance of sparse matrix multiplication and similar low compute-to-byte operations. These operations have low potential for parallelism, and instead will focus on improved techniques for element-wise operations.