Title: AI, Innovation and Global Transformation: Interdisciplinary Perspectives on Technology, Business and Society
Editors: Dr. J. Preetha, and Dr. Siddhartha Mehrotra
ISBN: 978-81-69857-64-2
Chapter: 11
DOI: https://doi.org/10.59646/809/11
Author: Dr. P. Deepa
Abstract
Autonomous digital ecosystems increasingly demand low-latency, privacy-preserving, and energy-resilient inference directly at the extreme edge. Traditional cloud-centric Machine Learning (ML) architectures suffer from severe bandwidth saturation, non-deterministic latency, and critical data vulnerability. Tiny Machine Learning (TinyML) addresses these limitations by enabling deep learning inference on resource-constrained microcontrollers ( SRAM,
Flash) operating within sub-milliwatt power envelopes. This chapter introduces an Adaptive Quantized-Sparse Neural Architecture Search (AQS-NAS) framework engineered specifically for heterogeneous, ultra-low-power edge nodes. By synthesizing dynamic mixed-precision quantization (
to
), structured channel pruning, and memory-aware graph scheduling, the framework maximizes task accuracy while strictly adhering to hardware constraints. Experimental evaluations across diverse edge benchmarks demonstrate that AQS-NAS reduces model memory footprint by up to 88.4%, lowers inference energy consumption by 76.2%, and maintains accuracy degradation within 1.2% across audio anomaly detection, and industrial predictive maintenance workloads.
Keywords: Edge Intelligence, TinyML, Mixed-Precision Quantization, Neural Architecture Search, Ultra-Low Power Microcontrollers, Autonomous Edge Ecosystems