Embedded Machine Learning Framework for Real-Time Prediction of User Information Needs in Intelligent Systems

Abstract

Embedded computing combined with machine learning has offered new opportunities of intelligent and context-aware systems able to make real-time decisions at the network edge. The paper presents a lightweight embedded machine learning system that can be used to make predictions on the information required by users dynamically in smart information retrieval scenarios. The suggested architecture incorporates enhanced neural designs into a Raspberry Pi 5-based edge device, making it be locally assertive to a cloud connexion. The system consists of a predictive core that is a compressed and quantised neural network, which has been trained with anonymised data of user interaction and search behaviour. The systematic optimization to 93 % prediction accuracy, 200 ms mean inference latency, and under 5 W power consumption allows real-time user intent prediction to be successfully run on the resource constrained hardware, which offers a basis to privacy preserving, low latency intelligent systems. It proves that it is practically feasible to introduce adaptive learning directly into edge devices, opening the way to the next generation of cognitive interfaces in which seamless interaction between perceptions, reasoning and individual user support is built into practical application.