Ilmu Komputer & AI editorial
Open AccessOA2024
Tiny Machine Learning: Progress and Futures [Feature]
TinyML enables deep learning on billions of IoT devices by co-designing algorithms and system stacks to overcome severe hardware constraints. This review surveys progress, introduces MCUNet for ImageNet-scale applications, and extends to on-device training.
Ji Lin; Li-Geng Zhu; Wei-Ming Chen; Wei-Chen Wang; Song Han· IEEE Circuits and Systems Magazine· 2024· DOI 10.1109/MCAS.2023.3302182
The core problem
Tiny machine learning (TinyML) is a new frontier of machine learning. By squeezing deep learning models into billions of IoT devices and microcontrollers (MCUs), we expand the scope of applications and enable ubiquitous intelligence. However, TinyML is challenging due to the hardware constraints: the tiny memory resource is difficult to hold deep learning models designed for cloud and mobile platforms. There is also limited compiler and inference engine support for bare-metal devices. Therefore, we need to co-design the algorithm and system stack to enable TinyML. This review discusses the definition, challenges, and applications of TinyML, surveys recent progress, introduces MCUNet, extends to on-device training, and presents future directions.
Innovation
The authors adopt a systematic review approach, examining the landscape of TinyML from hardware constraints to algorithmic solutions. They emphasize the need for co-designing algorithms and system stacks to overcome the memory and compute limitations of MCUs. The review covers recent advances in deep learning on MCUs, including model compression, efficient architectures, and specialized inference engines. A key contribution is the introduction of MCUNet, a system-algorithm co-design framework that enables ImageNet-scale AI applications on IoT devices. The methodology extends from inference to training, introducing tiny on-device training techniques that allow models to adapt on resource-constrained devices.
Introduction
Tiny machine learning (TinyML) is a new frontier of machine learning. By squeezing deep learning models into billions of IoT devices and microcontrollers (MCUs), we expand the scope of applications and enable ubiquitous intelligence. However, TinyML is challenging due to the hardware constraints: the tiny memory resource is difficult to hold deep learning models designed for cloud and mobile platforms. There is also limited compiler and inference engine support for bare-metal devices. Therefore, we need to co-design the algorithm and system stack to enable TinyML. This review discusses the definition, challenges, and applications of TinyML, surveys recent progress, introduces MCUNet, extends to on-device training, and presents future directions.
Why it matters
The authors analyze the challenges of TinyML, noting that the tiny memory resource is difficult to hold deep learning models designed for cloud and mobile platforms, and that there is limited compiler and inference engine support for bare-metal devices. They argue that co-designing the algorithm and system stack is essential to enable TinyML. The discussion extends to future directions, emphasizing that today’s “large” model might become tomorrow’s “tiny” model, and that the scope of TinyML should evolve and adapt over time. This perspective underscores the dynamic nature of the field and the need for continued innovation in both algorithms and systems.
Who should read this
Pembaca Research Digest dan praktisi rumpun ini.
Opening member content…