Reproduce “TS-Memory: Plug-and-Play Memory for Time Series Foundation Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sisuo Lyu, Siru Zhong, Tiegang Chen, Weilin Ruan, Qingxiang Liu, Taiqiang Lv, Qingsong Wen, Raymond Chi-Wing Wong and 1 more
TS-Memory adapts frozen time-series foundation models to domain shifts by distilling retrieval-based guidance into a lightweight PlugMem module. During training, retrieved historical analogues teach forecast corrections; at inference, the backbone and module run without database search. On ChronosBolt ETTm1, MSE fell from 0.421 to 0.381 and MAE from 0.383 to 0.371, within eight-benchmark long-horizon tests.
Key insights
- The adaptive knowledge provided by online retrieval can be effectively compiled and internalized offline into a compact parametric module, enabling robust adaptation without runtime search overhead.
- A confidence-gated distillation mechanism, which filters unreliable teacher signals based on the teacher's performance advantage and retrieval confidence, is crucial for effective and non-destructive knowledge transfer.
- Aligning retrieved candidates via channel-wise mean shifts and training PlugMem to learn incremental median corrections (Δ-alignment) allows it to capture domain-specific adaptive improvements relative to the frozen backbone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.