Online Program

Compute and Memory Efficient Machine Learning Classifier for Resource Constraint Devices (309611)

Keywords: Machine Learning, Optimization, Quantization, Tree Data Structure

The Machine Learning (ML) Classifiers are highly compute and memory intensive. Hence often data is collected from the resource constraint Edge devices, such as a health monitoring hardware near the equipment or processor board in a robot or inside an autonomous vehicle etc. and sent to centralized compute centers/ Data centers consisting of high-power servers, to train a ML Classifiers. The trained model (Inference engine) is deployed at the Edge. There are several problems in this approach. Firstly, individual Edge device may vary in their models/ makes, sensor accuracies etc., hence deploying a general-purpose Inference engine do not provide optimal performance. Secondly, any change in Edge device require retraining of the model. Thus, mass deployment and maintenance of Inference engines on Edge devices running under different environmental and operational conditions is a Challenge! In our paper, we addressed this challenge and propose a ML Classifier based on an Optimized Quantization technique and an Efficient Tree Data Structure (Patent applied), requiring ~100x less compute resources, ~1000x less memory, and ~5x faster training time, compared to a traditional ML Classifier.