LITEWAY: LIghtweight HAR via Temporal Efficient highWAY

Wearable human activity recognition (HAR) is often limited by the high computational and energy demands of deep learning models, particularly those relying on recurrent architectures like LSTMs or GRUs. These recurrent models are inherently sequential, which hinders parallel processing and increases inference latency on resource-constrained edge devices. The authors seek to develop a fully convolutional, lightweight framework that maintains high recognition accuracy while minimizing parameter counts and energy usage.

LITEWAY introduces a fully convolutional architecture that replaces traditional recurrent temporal modeling with a novel Structured Convolutional Temporal Modeling (SCTM) module. The framework consists of three main parts:

1. A feature extraction backbone that uses residual and depthwise-separable convolutions to downsample and refine sensor data.
2. The SCTM module, which captures long-range temporal dependencies using shared projections and gated pathways without the need for recurrent hidden states.
3. A lightweight classification head that uses attention-based pooling to aggregate features before a final linear classification layer.

The authors provide two versions: a 'Light' variant for maximum efficiency and a 'Full' variant for higher representational capacity. Both are evaluated across 16 diverse HAR datasets using a subject-independent protocol.

LITEWAY demonstrates that fully convolutional designs can outperform traditional CNN-RNN hybrids in both efficiency and accuracy. Across 16 datasets, LITEWAY variants consistently achieved competitive macro F1 scores while reducing model size by up to 9.52x and energy consumption by up to 3.14x compared to existing lightweight baselines like TinierHAR and MLP-HAR. The results indicate that LITEWAY effectively balances the trade-off between model complexity and recognition performance, proving that complex recurrent structures are not strictly necessary for high-quality temporal modeling in wearable HAR.