SLICED PROBABILITY METRICS FOR NEXT GENERATION MACHINE LEARNING
Abstract: Despite the impressive recent advances in machine learning (ML), and especially in deep learning, our current technology still faces significant hurdles to reach its full realization for tackling today's critical challenges. Deep learning methods remain greedy, opaque, brittle against adversarial attacks, and narrow in task generalization. In this talk, I will focus on overcoming some of the current deficiencies in deep learning using the new family of sliced probability metrics we recently proposed, including sliced-Wasserstein and sliced-Cramér distances. I will present recent results on overcoming catastrophic forgetting in continual learning and our recent works on generative modeling and domain adaptation to learn with zero or few labels. I will show the effectiveness of the sliced probability metrics using both theoretical and experimental results. Finally, I will conclude my talk with a remark on the connection between `generalized slicing' and deep neural networks, which provides new insights into neural networks' inner workings and provides novel future research directions.