Machine learning classification in resource constrained environments

Abstract

KEYWORDS: Machine Learning; Internet of Things; Resource Constrained Environments; Classification; Offline Learning; Online learning; Binary Class Classification; Multiclass Classification This thesis investigates the process of classification of data in extremely resource constrained environments. These are devices with a very small sized memory, modest computational unit, and finite sources of energy. The work is significant as Internet of Things’ (IoT) deployments today increasingly rely on learning algorithms to analyse collected data, draw conclusions, and take decisions on the further course of action. The norm is to deploy such decision making algorithms on the cloud and have the IoT nodes interact with the cloud on a regular basis. While this is effective for a number of applications, it is rather wasteful in terms of energy spent in communication, flooding the network, and temporal latency. In this thesis, we develop a set of techniques that facilitate classification within the resource constrained environments of IoT nodes. Two approaches are proposed for doing this. A batch learning approach that trains offline and comprising selecting a small number of prototypes/representative data points from a large data set. The selected prototypes are deployed on the IoT nodes. The prototypes are so selected that they appropriately represent the entire dataset and correctly classify new incoming data. The proposed technique is validated using publicly available standard datasets and compared with some state-of-the-art classification techniques. A prototypical implementation of the proposed technique further validates its efficacy.