Continual Learning

I led the development of Continual Learning at Scale project at Facebook AI. The goal of the project was to build SoTA continuous learning methods that optimizes trade-off between (i) forgetting vs intransigence, (ii) performance vs back-compatibility and (iii) accuracy vs performance. We developed CLAP (Continuous Learning as a Platform), a modular platform that can be used on top of PyTorch to enable SoTA continuous learning techniques including continuous samplers (importance and cluster based sampling), prompt-based continual learning, memory rehearsal techniques and self-supervision (to enable adapting to real-time streaming data as opposed to annotated data which take a week or more for data to stream through).