Videos

Efficient Tensor Representation for Deep Learning with TensorLy and PyTorch

Presenter
May 19, 2021
Abstract
Jean Kossaifi - Nvidia Corporation The data we manipulate in modern deep learning is inherently multi-dimensional. Preserving and leveraging that structure is crucial for good learning. Yet, this topological structure is typically discarded by existing models. By preserving and leveraging this structure using tensor methods, we can obtain better representations and enable better learning. This is particularly crucial when learning from spatiotemporal data or from structured data such as MRI. In this presentation, I will give an overview of tensor methods for deep learning for improved performance or speed, model compression and robustness. I will also cover practical implementation in PyTorch using TensorLy-Torch and show how to improve ResNet models for video based classification on the Kinetics dataset and for large-scale image classification on the ImageNet dataset. Biography: Dr. Jean Kossaifi is a Senior Research Scientist at NVIDIA. His current focus is tensor methods for machine learning. Particularly, efficient combination of these methods with deep learning to develop better models that are memory and computation efficient, while being more robust to noise, random or adversarial, as well as domain shift. He is the creator of TensorLy, a high-level API for tensor methods and deep tensorized neural networks in Python, designed to make tensor learning simple and accessible. He has also worked extensively on face analysis and facial affect estimation in naturalistic conditions, a field which bridges the gap between computer vision and machine learning. Prior to joining NVIDIA, Jean worked at the Samsung AI Center in Cambridge. He received his PhD and MSc from Imperial College London, where he worked with Prof. Maja Pantic. He also holds a French Engineering Diploma / MSc in Applied Mathematics, Computing and Finance and obtained a BSc in advanced mathematics in parallel.