I am a pediatric cardiologist and have cared for children with heart disease for the past three decades. In addition, I have an educational background in business and finance as well as healthcare administration and global health – I gained a Masters Degree in Public Health from UCLA and taught Global Health there after I completed the program.
“A mighty flame followeth a tiny spark.”
Dante Alighieri, Italian philosopher and writer
MCUNet is an exciting concept of microcontroller units (MCU) enabling machine learning on tiny IoT devices (tiny machine learning, or TinyML) from the research labs at MIT, National Taiwan University, and MIT-IBM Watson AI Lab.
It is a system-algorithm co-design framework that optimizes a neural architecture and an inference engine in the same loop: the MCUNet framework couples an efficient neural architecture search algorithm (TinyNAS) that optimizes the search space with the lightweight inference engine (TinyEngine) that directs resource management similar to an operating system.
In short, the TinyEngine provides the essential code to run the customized neural network of the TinyNAS. With these key two tiny components, ImageNet scale inference can be put on microcontrollers even with its very limited memory and processing power, and MCUNet is the first to achieve >70% ImageNet top1 accuracy with deep learning models. This is quite an engineering feat as the memory of a microcontroller is several orders of magnitude less than that of a cell phone and microcontrollers do not have operating systems nor much processing power.
The IoT with its 250 billion devices can benefit from this capability with its low-cost and low-energy microcontrollers to possess always-on machine learning. There is an obvious high degree of relevance for medical devices (especially wearable ones) in the context of personalized healthcare and chronic disease management for diseases (such as heart failure, diabetes, and chronic lung disease).
The biomedical data from these medical devices can then be analyzed locally without a central mechanism and the data may be more secure as there is no longer the need to send data to the cloud. In addition, the carbon footprint of a much smaller AI device is attractive (“green AI”).
This availability of a local small neural network for biomedical data heralds an exciting new era: it is akin to the development of a peripheral nervous system that has the capability to both process its signals locally as well as transmit its information centrally if it needs to.
The full study can be read here