Dynamic Bit Width Quantization Algorithms for Ultra Low Power Edge Intelligence

Abstract

The rapid expansion that comes with IoT ecosystems necessitates the use of deep learning algorithms in hardware nodes. However, the current fixed-point precision technique is computationally intensive and cannot be applied in constrained nodes. This paper offers solutions to these problems by introducing an adaptive approach that is designed specifically for precision scaling in localized systems. The technique uses runtime analysis of hardware constraints and data complexities to vary precision states during the algorithm's execution. The technique lowers precision states to ultra-low bit states in cases where there is a known input while raising precision states only during complex classification processes. Tests were conducted using specialized hardware simulators to measure processing delay, energy consumption, and accuracy of classifications. These experimental findings have revealed a reduction of up to 42.5% in terms of total energy usage in the system and a reduction of up to 60.1% regarding the memory footprint when compared to static bit-width networks. Most importantly, all these improvements were obtained while maintaining a baseline level of 94.8% accuracy for image classification tasks. Statistical analysis confirms that the proposed adaptive bit-width selection process manages to provide a balance between computational accuracy and energy preservation in physical systems. The current research proposes an effective way to implement deep learning processes on energy-constrained devices without constant access to a networked environment.