Tiny ML will move AI and IOT forward

#8 of my Tech Predictions for 2022

Two of the hottest technology areas are about to merge: Machine Learning and edge-based IoT (Internet of Things). This will create an explosion in use cases, by bringing intelligent, cheaper, less energy-consuming and more secure solutions into the world.

Machine learning is already exploding. New models, new algorithms, and new applications are popping up every day. IoT has also developed rapidly in the last few years. It is estimated that there are 250 billion microcontrollers embedded in devices in the world today. For example, in TVs changing channels and sound depend on signals from the remote control. And you find them in everything from office machines, cars, drones, medical devices, and home appliances. McKinsey researchers predict IoT will have a potential economic impact of US $5.5 trillion — $12.6 trillion by 2030, identifying manufacturing as the largest area (US $1.2–3.7 trillion)

But all these sensors have one thing in common: They generate lots of data at the point where they are placed and these data are used in Machine Learning applications. So data needs to be sent back and forth which slows things down — not so smart if the sensor has to decide whether to stop your car or not.

So, what if we could shrink deep learning networks to fit on tiny hardware and bring together Artificial Intelligence and intelligent devices? This is exactly what Tiny ML is all about. Tiny Machine Learning (TinyML) is a new discipline happening at the crossroads of machine learning (ML) and embedded systems that allows you to run ML models on low-power microcontrollers performing on-device analytics of sensor data at extremely low power.

Lots of great use cases

I know it sounds a bit hairy, but it´s a very big thing with a multitude of use cases.

Let me mention a few examples:

Always-on sound detection would become much better and could be integrated into things like elevators, door locks, and in-store kiosks, being used to detect whether a voice command has been heard (third floor, please).

The TinyML equipped elevator uses face recognition and voice activation

Forest fire detection is a growing problem, hard to solve. But Radio Controlled planes could prove to be a solution, flying autonomously while detecting and reporting wildfires using a normal onboard camera, a satellite modem, and TinyML doing image processing onboard.

-based forest fire detection

Range anxiety is another growing problem when more and more get electric cars. TinyML delivers much more trustworthy fuel consumption predictions by using a multitude of data collected while driving combined with clever fuel consumption machine learning models based on multiple factors such as distance, speed, temperatures inside and outside, AC, and other weather conditions.

Rainfall can be predicted by combining deep learning and different environmental parameters. With TinyML this can be used in devices that are offline and have to decide whether to perform an action based on whether it is going to rain or not. A great use case is the agricultural sector.

Things are heating up in the TinyML area

Apple recently acquired the TinyML startup Xnor.ai. They already use onboard chip algorithms in their Smartphones handling the billions of calculations happening every time you snap a photo.

The computer chip manufacturer ARM has joined forces with Neuton — a cloud-based platform, which allows the development of high-performance Machine Learning models 1000 times smaller than the ones we know today using Tensorflow or other AI frameworks.

The community even has its own TinyML Foundation started by people from Google and Qualcomm. They arrange Meetups around the world — even in Copenhagen.

TinyML provides better data security, lower power consumption, and standardization

A great upside of tinyML is increased data privacy and security since data (sound and images) are processed locally, never leaving the device. This also means extremely low power usage.

IoT devises today are typically custom-built for one particular purpose, which makes them quite expensive. But since the ML algorithm is software-based. The same hardware device could easily be reprogrammed and changed to adapt to a new use case. This enables the mass production of standard devices that can later be deployed for a wealth of different applications.

If this happens, it could very well be the IoT breakthrough we have been waiting for. Some even call it the iPhone moment of IoT. One thing is for sure: The world is about to get a whole lot smarter.

I predict that 2022 will be the year of convergence between edge-based tinyML and IoT.