Browse Teach Curate

Test-Time Adaptation for Time Series via Augmented Contrastive Clustering and Uncertainty-Aware Prototypes

Adapting a source-trained time-series classifier to distribution shift at inference using only unlabeled test data (ACCUP): an augmentation ensemble per input feeds uncertainty-aware class prototypes that down-weight low-confidence predictions, and an entropy-comparison scheme keeps only confident pseudo-labels for augmented contrastive clustering, tightening intra-class structure while limiting error accumulation from noisy self-training.

Data Text

Augmented Contrastive Clustering with Uncertainty-Aware Prototyping for Time Series Test Time

Test-time adaptation (TTA) updates a trained model on unlabeled test data to counter distribution shift at deployment; ACCUP targets the time-series setting, where such shift is common and prior TTA …