Time Series Classification on Edge with Lightweight Attention Networks

In recent years, wearable devices and sensing systems have become an integral part of the deployment of Artificial Intelligence (AI) on Edge. These intelligent edge devices perform first-level analytics to reduce data transfer, improve response latency, and preserve privacy. Thus, there is a growing focus on the design of machine learning models with high accuracy, and low resource footprint, suitable for a wide range of applications in the domains of healthcare, wellness, and lifestyle, among others. A majority of these tasks comprise the classification of time series signals collected from physiological and inertial sensors on commercially available wearable devices. A suitable solution for time series classification on edge would allow the use of multiple signal modalities on a broad spectrum of constrained devices. Unfortunately, the time series analytics for resource-limited platforms have not reached the same level of maturity as its full-fledged counterpart. For instance, attention-based networks, which form the state-of-the-art for standard time series classification, have no well-performing, generalized implementation for resource-constrained platforms. Towards this, we present a new architecture suited for time series classification using specialized lightweight attention-like blocks. These blocks, inspired by the attention mechanism and attention condensers, effectively learn time series features. We demonstrate the efficiency of the proposed model design and proof-of-concept implementation on a Human Activity Recognition dataset. Moreover, the generalizability of this architecture is highlighted by the reasonably good classification accuracy on various datasets from the UCR Time Series Classification Archive.