Olivia Weng and Alexander Redding had two invited presentations to describe our group’s latest and greatest reearch on fault-tolerant neural networks at the Fast Machine Learning for Science Conference. Alexander introduced Arbolta, a tool that bridges the gap between software and hardware fault injection by simulating faults in accelerators at the gate-level. Arbolta was born out of Alexander’s AMD internship under the supervision of Ian Colbert. Olivia presented on PrioriFI, a software fault injection tool that prioritizes flipping the most sensitive parameter bits in a neural network first. PrioriFI is a joint project with Nhan Tran from FermiLab.

There were many interesting presentations this year from across computer science, engineering, and physics. Our group enjoys attending FastML every year, meeting up with our collaborators across seas and sciences.