A small neural network trained on short time-series clips beats transformer-based models at finding anomalies — and runs faster doing it.
Researchers introduced PaAno, short for Patch-based representation learning for time-series Anomaly detection. The model chops time-series data into short temporal patches and runs each through a 1D convolutional neural network to produce a vector embedding. Training combines triplet loss and a pretext loss to make those embeddings carry real signal about normal behavior. At inference time, the model flags anomalies by comparing a time step's surrounding patch embeddings to those of normal patches from training data. Tested on the TSB-AD benchmark, PaAno outperformed existing methods — including transformer and foundation-model approaches — on both univariate and multivariate data, across point-wise and range-wise metrics.
The result is a pointed rebuke to the field's appetite for scale. Bigger architectures carry steep memory and compute costs that rule them out for real-time or edge deployments, and the paper argues those costs rarely buy meaningful accuracy gains under rigorous testing. PaAno's competitive performance on a standard benchmark suggests the community has been over-engineering the problem.
This is not the first time a lean method has embarrassed a large one on a time-series task — but systematic benchmarks like TSB-AD make it harder to dismiss as a cherry-picked comparison.