TS-Memory: Plug-and-Play Memory for Time Series Foundation Models
Researchers have developed TS-Memory, a novel plug-and-play memory adapter designed to enhance Time Series Foundation Models (TSFMs). This method addresses the challenges of adapting TSFMs to new domains by mitigating catastrophic forgetting and reducing inference latency. TS-Memory achieves this through a two-stage training process involving a kNN teacher and subsequent distillation into a lightweight adapter, enabling efficient, retrieval-free deployment. AI
IMPACT This new memory adapter could improve the adaptability and efficiency of time series forecasting models in real-world applications.