Can we predict future traffic, energy usage, or stock prices using historical data? While deep learning, particularly Transformer-based methods, has excelled in areas like computer vision and natural language processing, it has been found to underperform compared to simpler linear models in time series forecasting. This talk discusses recent hypotheses explaining why linear models remain competitive and presents empirical evidence supporting and challenging these ideas. Based on joint work with Nigus Teklehaymanot and Nhat Thanh Van Tran.
Elisha Dayag is a fourth year graduate student working with Jack Xin who is interested in all applications of deep learning, but especially computer vision and scientific machine learning.Â