5.5.22-Efficient-AI-OmniML

DI WU: All right. Hello, everyone. My name is Di Wu. I'm the co-founder and CEO of OmniML together with Professor Song Han from MIT. You'll hear more from Song this afternoon in his talk about TinyML and also the great technology, the great research that is behind our company.

So at OmniML, we focus on the edge AI, or empowering AI everywhere, which is common-- something believed by everyone that is the future of the AI in all the kinds of verticals, markets, and use cases. However, in the current industry, AI is still very, very hard, especially focusing a lot of the edge cases, right? The existing models are too big and too inefficient.

And also, there are so many diverse hardware platforms to optimize or to deploy these models are. Like, to learn all these hardware complexities, to learn about all these platforms, are really, really difficult. On top of that, the talents is also constrained. The people-- if you want to work on edge, if you want to work on AI for all these workloads, you really need people with both the ML knowledge as well as the hardware knowledge. And these people are very-- is extremely difficult to find.

And we believe one of the fundamental problems in edge AI is actually the mismatch between the models and then the hardware. Think about all these machine learning models have been evolving unconstrained in the past decade or so without any considerations of the hardware platform you are deploying these models to, where people are used to unconstrained resources from GPUs, from data centers.

However, once people want to bring these algorithms into the device, they're met with problems, right? The models are not designed for the device. And they really need to be optimized, need to be tuned, for it to be able to run in production.

And that process is painfully manual at this point, involving continuous iterations between the model designer, the ML engineers, as well as the hardware people, the hardware engineers who optimize and deploy these models on the device. So that leads to slow adoption, right? That leads to a long production cycle, R&D cycle. And that leads to less revenue, less volume.

And at OmniML, we are creating a software or a software service to help people make the model match the hardware all the way from the beginning, all the way from the model design phase. We insert ourself into the customer's workflow, and then we provide our customer, the ML engineers, capabilities of optimizing their model, focusing on all kinds of areas of constraints from hardware, from deployment.

And the central of that software is-- the technology is neural architecture search. These are hardware-aware, quantization-aware, making our model design process aware of all the constraints that we met in the deployment process.

And that software really unlocks the true potential of running AI everywhere on all the devices, from complex pose detection on mobile phones-- we're demonstrating real-time Qualcomm 855 chips-- to auto-driving scenarios, right? We are fusing six camera streams, doing 3D detection using an embedded GPU platform that is energy efficient, that can really runs in the card with batteries, with onboard power.

And also, we're making cameras smarter without updating the existing hardware, just a firmware update. On existing hardware, we can unlock potentials to run a lot of the computations closely on the edge, improving the calls, improving the privacy. And this technology even enables us to run complex computer vision algorithms on tiny microcontrollers with hundreds of kilobytes of memory.

And one of our early customers is Wyze. They're a user of our software, and they are a home security camera company. And their developers are using our software to improve their own model. And that unlocks tremendous amount of cost savings as well as improvement in their user experience.

And on top of that, we are also working with-- we've talked to more than 70 customers, we're working with multiple POCs, and then we're working with two of the major EV companies. Out there, we're working with their [INAUDIBLE] team, all of that in less than a year of in business. Just shows the attractions and the needs for this technology.

And for that, we're looking for partners. We're looking for customers in all the spaces, from smart cameras to manufacturing to robotics to personal health care to self-driving, autonomous vehicles, and love to talk to you. And come talk to me-- the exhibit-- to learn more. Thank you very much.

[APPLAUSE]