Workshop: FTXS: Workshop for Faults, Trustworthiness, and eXplainability for AI Systems at Scale
Authors: Luis Carlos N. Todd and Esteban Meneses (National High Technology Center, Costa Rica Institute of Technology); Elvis Rojas (National University of Costa Rica); and Florina Ciorba (University of Basel)
Abstract: High Performance Computing platforms have nurtured the development of deep learning (DL) models, particularly providing the massive computing power needed for training those models. However, the alarming growth rate of DL models has made their optimization, to reduce memory size and energy consumption, a priority. In particular, precision reduction techniques have been developed to tackle those challenges. The available literature provides a good description of how precision reduction affects the performance deep learning models. However, there is a knowledge gap of how sensitive precision reduction techniques are to silent data (SDCs), that can lead to unsuspected errors. SDCs are becoming more prevalent, for that reason, we set out to investigate how deep learning models, optimized with precision reduction techniques, react to bit flips caused by silent data corruption. Our results show that those optimized models are, in most cases, surprisingly resilient to bit flips compared to original model.
Back to FTXS: Workshop for Faults, Trustworthiness, and eXplainability for AI Systems at Scale Archive Listing
Back to Full Workshop Archive Listing