Non-autoregressive Conditional Diffusion Models for Time Series Prediction
Jul 24, 2023
Speakers
Lifeng Shen
Speaker · 0 followers
James T. Kwok
Speaker · 0 followers
About
Denoising diffusion models have led to breakthroughs in synthesizing images. However, it is still an open question of how to adapt their strong modeling ability to real-world time series prediction. Compared to main-stream deep prediction methods (e.g., Transformers), existing time series diffusion models' performance is still unsatisfactory. This work proposes a new denoising diffusion model for time series prediction. The model is non-autoregressive and achieves high-quality prediction by developing two conditioning mechanisms during training: future mixup and autoregressive initialization. The former allows some parts of ground truth future predictions in conditioning, while the latter helps capture basic time series patterns (e.g., short-term trends). Based on the combination of these two parts, the proposed method thus has a better denoising process to generate future predictions. We demonstrate experimentally that the proposed non-autoregressive denoising diffusion model is the new state-of-the-art diffusion method for time series prediction and achieves better or comparative results than recent strong baselines, e.g., Transformers and FiLM.Denoising diffusion models have led to breakthroughs in synthesizing images. However, it is still an open question of how to adapt their strong modeling ability to real-world time series prediction. Compared to main-stream deep prediction methods (e.g., Transformers), existing time series diffusion models' performance is still unsatisfactory. This work proposes a new denoising diffusion model for time series prediction. The model is non-autoregressive and achieves high-quality prediction by deve…
Like the format? Trust SlidesLive to capture your next event!
Professional recording and live streaming, delivered globally.
Sharing
Recommended Videos
Presentations on similar topic, category or speaker
FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU
Watch later
Ying Sheng, …
Provably Correct Physics-Informed Neural Networks
Watch later
BNN-DP: Robustness Certification of Bayesian Neural Networks via Dynamic Programming
Watch later
Steven Adams, …
