Sequential Data Classification for Resource-constrained Devices

We study the problem of fast and efficient classification of sequential data (such as time-series) on tiny devices, which is critical for various IoT related applications like audio keyword detection or gesture detection. Deploying sequential data classification modules on tiny devices is challenging as predictions over sliding windows of data need to be invoked continuously at a high frequency. Each of these predictors themselves are expensive as they evaluate large models over long windows of data. In this paper, we address this challenge by exploiting the following two observations about classification tasks arising in typical IoT related applications: (a) the "signature" of a particular class (e.g. an audio keyword) typically occupies a small fraction of the overall data, and (b) class signatures tend to discernible early-on in the data. We propose a method that exploits these observations by using a multiple instance learning formulation along with an early prediction technique to learn a model that can achieve better accuracy compared to baseline models, while reducing the computation by a large fraction. For instance, on an audio keyword detection benchmark our model improves standard LSTM model's accuracy by up to 1.5\% while decreasing the computation cost by more than 60\%. This enables us to deploy such models for continuous real-time prediction on a small device such as Raspberry Pi0, a task that the baseline LSTM could not achieve. Finally, we also provide an analysis of our multiple instance learning algorithm in a simple setting and show that the proposed algorithm can efficiently converge to the global optima, one of the first such result in this domain.