11.8.1 Automating Tiny Neural Network Design with MCU Deploy-ability in the Loop
Tiny Machine Learning (TinyML) is a growing, widely popular community focusing on the deployment of Deep Learning (DL) models on microcontrollers (MCUs). To run a trained DL model on an MCU, developers must have the necessary skills to handcraft network topologies and associated hyperparameters to fit a wide range of hardware requirements including operating frequency, embedded SRAM and embedded Flash memory along with the corresponding power consumption requirements.
Unfortunately, a hand-crafted design methodology poses multiple challenges: 1) AI and embedded developers exhibit different orthogonal skills, which do not meet each other during the development of AI applications until their validation in an operational environment 2) Tools for automated network design often assume virtually unlimited resources (typically deep networks are trained on cloud- or GPU-based systems) 3) The time-to-market from conception to realization of an AI system is usually quite long. Consequently, mass market adoption of AI technologies at the deep edge is jeopardized.
Our solution is based on Sequential Model Based Optimization (SMBO) – aka Bayesian Optimization (BO) – that is the standard methodology for Automated Machine Learning (AutoML) and Neural Architecture Search (NAS). Although AutoML and NAS are successfully applied on large GPU/Cloud platforms (i.e., some AutoML/NAS tools are commercialized by Google, Amazon and Microsoft), their application is still an issue in the case of tiny devices, such as MCUs. Our approach, instead, includes “deployability” constraints – related to the hardware resources of the MCUs – into the hyperparameter optimization process, leading to this new “AutoTinyML” perspective.
This talk will present our approach, along with its pros and cons with respect to multi-objective optimization (usually adopted to reduce resource usage on cloud). A set of relevant results will be presented and discussed, providing an overview of the next open challenges and perspectives in the AutoTinyML field.