Weightless is a novel scheme for lossy weight encoding. The encoding is based on the Bloomier filter, a probabilistic data structure that can save space at the expense of introducing random errors. Leveraging the ability of neural networks to tolerate these imperfections and by re-training around the errors, Weightless compresses weights by up to 496× without loss of model accuracy. This results in up to a 1.51× improvement over the state-of-the-art.
Ares is an algorithmic fault injection framework to facilitate an academic understanding of fault tolerance in deep learning models. Ares enable fault tolerance studies across models, structures, and different datatypes. The code is available here.
It has been an amazing 6 years. I've been fortunate enough to have worked with two amazing advisors and collaborated on building simulators, benchmarks, chip design, machine learning, and robotics! I'm excited to start the next phase of my career working with the AI Infrastructure Research team and Kim Hazelwood at Facebook in July!