Focus on privacy and new features will make TinyML even more powerful

By: STACEY HIGGINBOTHAM

Privacy and smart features that are not dependent on applications will likely drive the adoption of machine learning (ML) on constrained edge devices forward. That was the message Edge Impulse CEO Zach Shelby and I tried to convey when we sat on a virtual panel at the TinyML Summit this week.

In our discussion on TinyML in the smart home, we focused on use cases and what it will take for companies to adopt machine learning on constrained devices. First, some vocabulary, TinyML is focused on providing machine learning or inference on microcontrollers. Lately, this definition has been expanded a bit to include all ARM-based processors, such as the Cortex series processors used in smartphones. So not all of the use cases we mentioned can be run on battery-powered microcontrollers.

The benefits of incorporating machine learning into embedded devices are enormous.

No matter how you define the Tiny in TinyML, it's already an important trend as consumers worry about privacy and industry tries to build devices that use less power and respond faster and faster. Shelby said he's seeing a lot of demand for TinyML in the industrial sector, as well as from white goods manufacturers, for predictive maintenance with TinyML. In these examples, machines might have sound or vibration sensors that learn the sounds or vibrations from normal machinery and send out alerts when either of those changes.

With machine learning on the chip, each sensor can be personalized to fit how the device operates in a specific environment. For example, in a smart home, a washing machine or refrigerator with smart sensors can send a signal before the motor or compressor breaks down. TinyML might also be used to calculate the weight of the clothes in the washing machine, etc., and adjust the water level accordingly.

Shelby noted that customers also want TinyML to do fine-grained location tracking of goods in a warehouse, office or home. Other use cases for on-device machine learning involve new services in wearables, health care and smart homes. Since privacy is a key advantage of on-device processing, health care is a good market because laws like HIPPA can make doctors or hospitals nervous about sharing data through connected devices.

Wearables that use machine learning to process sleep data or heart rate data locally might let users track their health without uploading it to Apple or Google servers. A closed-loop insulin pump could use ML to measure glucose and signal the release of insulin, all without an always-on internet connection. This could keep medical data truly private and make it more secure when there's no internet connection.

In the home, using tools like radar can help devices figure out how many people are in a room or how close one device is to another. This can provide the necessary context for a smart home without relying on a smartphone or camera. And yes, we also talked about creating custom wake words or off-the-cloud wake words, which will let devices respond to a limited set of spoken commands. For example, if a lamp has a microphone and chip that can interpret a few phrases, such as "turn on the lights" or "dim the lights," it doesn't have to be connected to the internet.

I tried to break out a few areas where I think TinyML could have a big impact. I thought about privacy-focused security systems; closed-loop sensor/actuator systems like insulin pumps or even NVAC systems that measure air quality; single-purpose devices that don’t need to be connected, but just need some intelligence, like a mattress that tracks a baby’s breathing and sounds an alarm if it stops; and devices that need super-fast response times, like sensors on motors for preventative maintenance, etc.

To this list, Shelby also adds customized interfaces. He also gives the example of using custom wake words to activate devices or new interfaces that can include gesture or gaze detection. By using on-device ML, such interfaces will also protect the user's privacy. We haven't mentioned the security benefits of using local machine learning, but they are considerable. After all, if you don't connect your device to the internet, your attack surface is much smaller. This benefit goes hand in hand with privacy.

At the end of the talk, we provided some resources for those who might want to play around with TinyML or see it in action. An audience member shared his enthusiast equipment, Shelby mentioned a free Coursera course on embedded machine learning, and Edge Impulse is working with ARM, Arduino, and the TinyML Foundation to help developers accelerate edge machine learning applications.