EAGER: Low-Energy Architectures for Machine Learning

点击翻译按钮获取中文摘要

Machine learning systems and classifiers will be part of future smart devices. Industrial internet-of-things (IIOT) and cyber-physical systems (CPS) will be equipped with real-time feature extraction and classification to provide feedback and/or warning signals in some cases. Smart medical devices can analyze signals and trigger therapy to improve human health. Security systems can analyze activity data and thwart planned attacks. Reducing energy consumption in these smart devices is critical for increasing battery life in portable applications. This proposal addresses techniques to reduce energy consumption in feature extraction and classification. The broader impacts will be in demonstrating a new approach for feature extraction and classification with significantly less energy consumption without degrading sensitivity and specificity, along with training and educating graduate and undergraduate students in related disciplines through laboratory and computational experiences.The proposed framework computes features and classifies the test data using a simple level-1 classifier that makes use of low precision. If the classification is successful, then the process terminates. Otherwise the level-2 classifier is invoked. The level-2 classifier makes use of higher precision for the feature extraction and classification; however, it reuses the low-precision results of the level-1 classifier. The process is repeated in an iterative manner until the test sample is classified with a high probability. The proposed approach differs from existing approaches in the sense that the classifier at a certain level is trained using only the training samples that do not contain the samples that were correctly classified in prior levels. The precision at the different levels of feature extraction and classification are the same for both training and test phases. This is expected to lead to higher classification accuracy. The features and classifiers are computed using approximate computing in an incremental manner. Other innovative aspects include: selection of classes of features that require less energy (e.g., time-domain vs. frequency-domain), ranking of these features using techniques such as minimally-redundant maximally-relevant (mRMR) and use of classifiers such as classification and regression tree (CART) or AdaBoost. Approximate computing of features and classifiers in an incremental manner will be investigated to reduce overall energy consumption while maintaining high sensitivity and specificity. Training of the P-Boost classifier and testing the classifier will be based on same precision; thus there is no disconnect between the precision of the classifiers used for training and testing. The proposed "holistic" approach is likely to result in significant savings in energy consumption compared to state-of-the-art machine learning systems.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.