Revisiting Transformer: Inversion More Effective, New SOTA for Real-World Prediction
The Transformer has shown strong capabilities in time series forecasting, capable of describing pairwise dependencies and extracting multi-level representations from sequences. However, researchers have also questioned the effectiveness of Transformer-based predictors. These predictors often embed multiple variables at the same timestamp into indistinguishable channels and attend to these time tokens to capture temporal dependencies. Considering … Read more