Reproduce “TRACE: Memory-Efficient Time Series Class-Incremental Learning via Temporal-Spectral Reconstruction” 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
Junhui Mao, Junru Zhang, Liming Ye, Cheng Peng, Haoran Shi, Xiaoli Tang, Baokun Wang, Bin Dou and 2 more
TRACE addresses continual time-series recognition by freezing a temporal-and-frequency feature extractor and adding an independent autoencoder for each new class. At prediction, the class model with the lowest reconstruction error wins. Its five-dataset macro-average final accuracy is 88.85% with 9.17 MiB retained state, versus TS-ACL at 84.61% with 525.68 MiB; old-class accuracy can still decline as new classes compete.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.