Improving Resource Efficiency of Deep Activity Recognition via Redundancy Reduction

Compression methods for deep learning have been recently used to port deep neural networks into resource-constrained devices - such as digital gloves and smartwatches - for human activity recognition (HAR). While the results have been in favor of utilizing compressed models, we envision that the current paradigm of long and fixed-size overlapping sliding windows that permeate the literature of HAR contributes negatively toward the goal of more resource-efficient systems, as it induces redundancies in memory and computation. In this work, we provide a different perspective by demonstrating that memory footprint, computational expense, and possibly energy consumption can be dramatically spared by modifying the architecture of the neural networks and their training. It is achieved by enabling non-overlapping short sliding windows and skipping fine-grained features in favor of rough ones on certain occasions, thus reducing the demand for more powerful hardware. Compared with the state-of-the-art, our method is able to achieve comparable performance far more efficiently in terms of resource use.