Tiny machine learning, often referred to as tinyML, is a field of study within Machine Learning (ML) and embedded systems that explores the types of models that users can run on small, low-powered devices. Typical ML models are run out of large data centers or clouds with clusters of central processing units (CPUs) and graphics processing units (GPUs), each of which could require power sources anywhere from 65 to 500 watts. In tinyML, models can instead be run on microcontrollers, computers that can be as small as a grain of rice that only consume around miliwatts or microwatts of power. By utilizing microcontrollers, tinyML is able to address space and power challenges in embedded AI by hosting machine learning models locally within a device, such as a smartphone, while consuming almost 1000 times less power.
While the tinyML industry is still in its infancy, there has been a lot of activity in the space within recent years. Since January 2020, over $26 million has been invested in tinyML, including from venture capital accelerators, early-stage investors and late-stage investors, according to an emerging spaces review by Pitchbook. The most common applications of tinyML technology are within the fields of audio analytics, pattern recognition and voice human machine interfaces, although proponents of tinyML hypothesize that predictive maintenance is likely to be one of the most common use cases with the highest impact.
