Reg-TuneV2: A Hardware-Aware and Multiobjective Regression-Based Fine-Tuning Approach for Deep Neural Networks on Embedded Platforms

Fine-tuning deep neural networks (DNNs) for deployment has traditionally relied on computationally intensive methods such as grid searches and neural architecture searches, which may not consider hardware-aware metrics. Moreover, it is essential to consider multiple objectives to develop a range of solutions for tiny machine learning hardware deployment with real-time latency and low power constraints. To address these problems, we propose Reg-TuneV2, a systematic approach to fine-tune DNNs for hardware deployment by considering multiple objectives, including accuracy, power, and latency contours. In addition, this approach uses metric learning to achieve smaller and better-suited configurations for deployment, achieving 90.5% accuracy with only 340 KB of memory for keyword spotting (KWS) on a field-programmable gate array. When compared to baselines for KWS and image classification on the Nvidia Jetson Nano 4-GB Software Development Kit, the proposed method achieves a 14.5$ \times $× and 101.8$ \times $× reduction in model size coupled with 2.5$ \times $× and 5.9$ \times $× better inference efficiency, respectively.