LITEWAY: LIghtweight HAR via Temporal Efficient highWAY

Summary

Wearable human activity recognition is constrained less by accuracy than by the computational and energy budget available on resource-limited devices. The authors observe that existing lightweight HAR approaches lean on recurrent architectures such as GRU and LSTM, which limit parallelism and increase inference latency on embedded targets. LITEWAY is proposed as a modality-agnostic, fully convolutional framework for multichannel sensor time series that replaces recurrent temporal modeling with structured convolutional decomposition.