In the paper, TinyML Meets IoT: A Comprehensive Survey, the authors noted that the rapid growth in miniaturisation of low-power embedded devices and advancement in the optimisation of machine learning (ML) algorithms have opened up a new prospect of the Internet of Things (IoT), tiny machine learning, which calls for implementing the ML algorithm within the IoT device.
Tiny Machine Learning, or TinyML, is typically used for automated tasks involving sensory data. As TinyML vendors continue to democratise Machine Learning (ML) at a rapid pace, ABI Research forecasts that TinyML Software-as-a-Service (SaaS) revenue will exceed US$220 million in 2022 and become an important component from 2025 onward.
While total revenue will be dominated by chipset sales, as TinyML device shipments continue to grow, the TinyML SaaS and professional service market have the potential to become a billion-dollar market by 2030.
The TinyML market has come a long way since ABI Research first analysed this market back in 2020. The TinyML Foundation, which gathers most of the main vendors in this space, has greatly expanded in recent years. And so have the applications of TinyML, with forest fire detection, shape detection, and seizure detection among some of the most spectacular use cases.
Given how central environmental sensors are to TinyML, the possibilities are extensive. David Lobina, artificial intelligence & machine learning research analyst at ABI Research, explains, “Any sensory data from an environment can probably have an ML model applied to that data.”
He listed out some of the most common applications including Word Spotting (the identification of keywords in text or utterances), Object Recognition (the detection of a person by a sensor), Object Counting (a sensor that counts the number of people inside a building), and Audio or Voice Detection, (as in the models that activate upon hearing ‘Hey, Google’).
Ambient sensing and audio processing remain the most common applications in TinyML, with sound architectures holding an almost 50% market share in 2022. Most of these applications employ either a microcontroller (MCU) or an Application-Specific Integrated Circuit (ASIC). The personal and work devices sector will be the largest increase soon.
With the myriad possibilities, there are also potential pitfalls, but for which, ABI Research believes there are well-identified solutions. “The physical constraints on TinyML devices are genuine. These devices favour small and compact ML models, which call for innovation at the software solutions level for specific use cases. And software providers will be the most active in the TinyML market,” says Lobina.
ABI Research recommends vendors concentrate on those applications that TinyML has a clear value proposition worked out before production.
Lobina says the role of software is crucial, and vendors must develop software tools to automate TinyML itself, a recursive process that necessitates employing TinyML applications to automate other TinyML applications.
“And finally, new technology will be required to bring about ever more sophisticated TinyML models. Neuromorphic computing and chips, along with the corresponding technique of Spiking Neural Networks, would bode well for the future,” he concludes.