Privacy and new functions will make TinyML big

Privacy and smart features that don’t depend on an app will likely drive the adoption of machine learning (ML) on constrained edge devices going forward. That was the message Zach Shelby, CEO of Edge Impulse, and I tried to convey when we sat on a virtual panel at the TinyML Summit this week.

During our chat about the general use of TinyML in the smart home, we focused on use cases and what it will take to get companies to adopt machine learning on constrained devices. First, some vocabulary. TinyML focuses on providing machine learning or inference on microcontrollers. The definition has been expanding a bit as of late to include all ARM-based processors, such as those in the Cortex family, which are used in smartphones. So not all of the use cases we mentioned could run on battery-powered microcontrollers.

But no matter how you define the tiny in TinyML, it’s already a significant trend as consumers worry about privacy and companies try to build devices that use less power and respond ever more quickly. Shelby said he’s seeing a lot of customer demand for TinyML in the industrial sector as well as from white goods manufacturers, for predictive maintenance involving TinyML. In these examples, machines might have sound or vibration sensors that learn the sounds or vibrations coming from a normal machine and send alerts when any of them change.

With actual learning on the chip, each sensor could become personalized to the ways in which the device runs in a particular environment. In a smart home, for example, a washing machine or fridge with sensors could send a signal before a motor or compressor breaks. TinyML might also be used to calculate something like the weight of clothing in the washer, so that the water levels can be adjusted accordingly.

Shelby noted that customers also want TinyML for the fine-grained location tracking of goods in warehouses, offices, or homes. Other use cases for on-device machine learning involve wearables, health care, and new services for the smart home. With privacy as a key benefit to on-device processing, health care is a great market because laws like HIPPA can make doctors or hospitals nervous about sharing data over connected devices.

Wearables that use machine learning to process sleep data or heart-rate data locally might let a user track their health without having to upload it to Apple or Google. A closed-loop insulin pump could use ML to measure glucose and signal the release of insulin, all without having an always-on internet connection. This could keep medical data truly private, and without an internet connection, render it more secure.

In the home, using tools such as radar could help devices figure out the number of people in a room or the proximity of one device to another. This could provide necessary context for a smart home without relying on a smartphone or a camera. And yes, we also talked about creating customized wake words or non-cloud-based wake words that would let a device respond to a limited set of spoken commands. For example, a light wouldn’t have to be connected if it had a microphone and chip capable of deciphering a few phrases, such as “Turn on the light” or “Dim the light.”

I tried to break out a few areas where I think TinyML could have an outsized influence. I came up with privacy-centric security systems; closed-loop sensor/actuation systems, such as an insulin pump or even an NVAC system that tracks air quality; single-purpose devices that don’t need connectivity, just some smarts, such as a mattress that could track a baby’s breathing and send an alert if it stops; and devices that need a super-fast response time, such as  asensor on a motor detecting a problem and stopping it before it breaks.

To that list Shelby added customized interfaces. He also gave an example of using custom wake words to activate devices or new interfaces that could include gestures or gaze detection. By using on-device ML, such interfaces would also preserve user privacy. Neither of us mentioned the security benefits of using local machine learning, but they are considerable. After all, if you don’t connect a device to the internet, you have a much smaller attack surface. That benefit goes hand in hand with privacy.

We ended the talk by providing some resources for those who might want to play with TinyML or see it in action. An audience member shared his fever-detection device and Shelby mentioned a free Coursera class on embedded machine learning Edge Impulse helped develop along with ARM, Arduino, and the TinyML Foundation. Check it out. I think TinyML is going to be big.