MACHINE LEARNING TECHNIQUES FOR HIGHLY ENERGY-EFFICIENT CIRCUITS

Abstract

Energy-quality (EQ) scaling in the context of machine learning is discussed. Three designs are proposed to evaluate the effects of EQ scaling in voice activity detection (VAD), as a specific test case of sensor nodes in the Internet of Things (IoT) which incorporate machine learning algorithms. An EQ-scalable VAD system is first presented. Knobs are co-optimized dynamically to extend the range of energy reduction and mitigate effects of under/overfitting. The importance of context-specific retraining is demonstrated. Next, an EQ-scalable self-learning VAD system is proposed. Labels are generated within the system, and training-time computations are simplified. Finally, a dual-mode ADC further extracts power savings under low signal activity. Clock activity can be inhibited when new samples fall within a range defined by programmable offset of complementary comparators. A low-cost offset calibration scheme is also introduced. Further system-level savings beyond local activity gating can thus be pursued.