No old samples, one autoencoder per class: TRACE cuts retained memory about 98% below TS-ACL
Synopsis
For behavioral classes that arrive over time, TRACE stores no historical samples and never expands a shared classifier; it feeds a frozen temporal-spectral representation to a small autoencoder per class and labels by reconstruction error. Across five multivariate time-series class-incremental benchmarks it reaches 88.85% five-dataset average final accuracy with 9.17 MiB of retained memory, about 98.26% less memory than TS-ACL, the strongest analytical baseline by average accuracy, and an approximately 33.43% relative accuracy gain on GrabMyo.
Interpretation
A time-series classifier can now take on new classes without keeping historical samples or repeatedly rewriting a shared decision boundary. Earlier analytical continual-learning methods froze the backbone but expanded feature dimensions sharply to avoid forgetting, paying for large covariance matrices that exceeded 500 MiB of retained memory in some configurations. On HAR, UWave, DailySports, WISDM and GrabMyo, TRACE reaches 88.85% five-dataset average final accuracy with 9.17 MiB retained, while TS-ACL reaches 84.61% with 525.68 MiB.
When a new class arrives, the system trains one independent small autoencoder for it, and inference asks which class autoencoder reconstructs the feature best. An autoencoder is a network that compresses an input and reconstructs it; here it models one class's feature distribution rather than serving as a shared classifier. The extractor is a four-block 1D CNN fed the raw signal concatenated with the magnitude spectrum of a single-sided real fast Fourier transform, frozen after initial-task training together with its standardization statistics; at the default 576-dimensional features, each new class's autoencoder costs roughly a few MiB in FP32.
The scoring function of every previously learned class stays fixed across later tasks, so old-class recognition is not rewritten as new classes join. Shared-model updates let old knowledge drift, whereas each learned class autoencoder is frozen after training and never accessed or updated again. On GrabMyo, TRACE shows a mean old-class accuracy decline of 1.24 percentage points, against 26.27 points for the generative-replay method GR.
Perspective
Evaluation covers five multivariate time-series benchmarks, HAR, UWave, DailySports, WISDM and GrabMyo, with two new classes per incremental task; results average five class orders and five random seeds with 95% confidence intervals.
The source notes that broader pretraining coverage generally improves recognition, but expanded and all-class pretraining use future-task labels and are therefore evaluated only as representation diagnostics; on UWave and WISDM, TRACE is competitive with KLDA and TS-ACL rather than ahead.