Abstract
Most neural networks (NNs) generated by neural architecture search (NAS) are discarded except for the final output to limit the memory usage on high performance computing (HPC) systems on which the search is performed. However, discarded NNs are vital for understanding the NAS structure’s evolution and reproducibility. We design a visual interactive tool for NN archaeology that explores the evolution of NAS structures, finds matching subsequences in the structures, and visualizes NN similarities across NAS outputs, including discarded NNs. We demonstrate the capabilities of our tool to discover and visualize matching subsequences on a dataset of NNs generated by NSGA-Net, a genetic NAS.