Self-Bidirectional Decoupled Distillation for Time Series Classification

Over the years, many deep learning algorithms have been developed for time series classification (TSC). A learning model's performance usually depends on the quality of the semantic information extracted from lower and higher levels within the representation hierarchy. Efficiently promoting mutual learning between higher and lower levels is vital to enhance the model's performance during model learning. To this end, we propose a self-bidirectional decoupled distillation (self-BiDecKD) method for TSC. Unlike most self-distillation algorithms that usually transfer the target-class knowledge from higher to lower levels, self-BiDecKD encourages the output of the output layer and the output of each lower level block to form a bidirectional decoupled knowledge distillation (KD) pair. The bidirectional decoupled KD promotes mutual learning between lower and higher level semantic information and extracts the knowledge hidden in the target and nontarget classes, helping self-BiDecKD capture rich representations from the data. Experimental results show that compared with a number of self-distillation algorithms, self-BiDecKD wins 35 out of 85 University of California, Riverside (UCR) 2018 datasets and achieves the smallest AVeraGe (AVG)_rank score, namely 3.2882. In particular, compared with a nonself-distillation baseline, self-BiDecKD results in 58/8/19 regarding 'win'/'tie'/'lose.'