NBEATS - Time-series forecasting using feed-forward networks
Boris Oreshkin
The author will talk about N-BEATS neural network architecture and its applications to time-series forecasting. He will describe the architecture design, motivation and cover extensive empirical evidence demonstrating its effectiveness in various scenarios such as zero-shot transfer across datasets and small data learning for electricity load forecasting.
About
Boris Oreshkin is currently a research scientist at Unity Technologies working on developing machine learning tools for AI assisted character animation in games. He has 15 years of experience in academic and industrial R&D covering areas of machine learning, signal processing, time series analysis, medical image processing and data science. While being an applied research scientist at Element AI, Boris developed the N-BEATS neural network architecture for univariate time-series forecasting in collaboration with Dmitri Carpov, Nicolas Chapados and Yoshua Bengio. His research interests include learning with limited labels, weakly supervised learning, multi-modal information processing, feed-forward and attention based architectures, information processing on graphs and modelling kinematics with machine learning.
