5.5.22-Efficient-AI-Pison

HENRY VALK: Hello. My name is Henry Valk, and I'm a data scientist at Pison Technology.

Pison creates technologies that use the body's electrophysiological signals to harness the power of human intent. The company was founded in 2016 by David Cipoletta and MIT alumnus Dexter Ang. And early in the company's history, we received support from the MIT Sandbox Innovation Fund. So in a lot of ways, Pison is an MIT startup.

At Pison, we're developing the next generation of human/machine interfaces by integrating gesture control technology into existing consumer and defense products. The future is moving towards human input devices that are natural and intuitive to use, that allow users to remain heads-up and hands-free, and that operate at the speed required by immersive computing platforms. Here are a few examples of the gestures our technology can support.

And notice that these gestures already have meanings that you understand, right? Over there, stop, good. This reduces the learning curve for adopting our gestural language, and ultimately reduces the cognitive load of interfacing with your devices.

So how does it work? Electrodes on the bottom of a wrist-worn device capture electrical potentials originating from the motor cortex as they pass through the nerves and muscles in the wrist and propagate to the surface of the skin. We then digitize that analog signal and transmit it over Bluetooth to a host device. From there, the digital electromyographic signal, EMG, is decoded into Pison's proprietary electroneurographic signal, ENG, which is able to differentiate between nerve and muscle signals.

And to add a bit more color on what our algorithms are doing with the Pison ENG signal, I'll describe what we call our L1 algorithm. This algorithm can be trained on a small amount of data, using just the compute on an average smartphone. It uses an outlier detection, based on the Mahalanobis distance, between the incoming ENG features and previously calculated feature distributions in order to reject noisy samples. We then use conventional machine learning methods to generate model inferences, and then Bayesian methods to aggregate those inferences into our final classification output.

As we push into the consumer space and begin to build a database of human intent on the scale of hundreds of millions of users, we'll be able to deploy our L2 models, which employ deep learning architectures. Similar to how Amazon leveraged proprietary hardware, data at scale, and cloud computation to rapidly improve natural language processing, and how Tesla leveraged cloud compute, data at scale, and proprietary hardware to rapidly improve autonomous driving, we hope to do the same thing with natural gesture language processing.

There are alternative solutions to solving gesture control. Computer vision is fine, as long as you're comfortable holding your hands in front of a camera that you're wearing on your face at all times, and as long as you never need to use your device at night. There are other wrist-worn solutions, but solutions that use blood oxygen levels or IMU alone that are capturing a mechanical signal, like the vibration from a tap, are unresponsive, slow, and unintuitive to use. They'll never be able to achieve the naturalness required by users or the speed required by immersive spatial computing platforms. Pison's ENG signal, because it's upstream of the hand, directly capturing intent from the motor cortex, we're actually able to detect gestures 30 milliseconds before your hand ever moves.

So let's take a look at one of our consumer use cases. This is our integration with the Google Pixel buds. As you can see, the user's able to control her music and the volume just with simple hand gestures. We can also map gestures performed in different orientations to different commands in order to substantially increase the size of our gesture vocabulary.

We've built confidence in our technology by shipping products to users with more at stake, namely the Air Force and the Air National Guard, users who need heads-up capabilities not to remain socially engaged, but to remain situationally aware. And while we're actively expanding our defense partnerships, we're also partnering with one of the world's largest chip manufacturers to embed our technology into any smartwatch, or even a custom, standalone gesture wearable. If you can think of a use case for hand gestures in your industry, we can build you a solution. So don't hesitate to reach out. Thank you.

[APPLAUSE]