Conceptual

Graph-Explanation-Guided Layer Fusion for Deep Neural Network Inference

A method that applies Graph Explanation Techniques (explainable-AI attributions over graph-structured models) to deep-neural-network layer fusion. When a fusion group is invalid because it exceeds an accelerator's on-chip buffer, graph explanation identifies the operations most responsible for the invalidity and a greedy tree-based algorithm recursively splits the group along them to minimize DRAM accesses, reducing inference data movement on convolutional networks.