We hear the hype about how fast AI is advancing and, whether its 5x faster than Moore’s Law, the reality is we need to make some adjustments. That was one of the points of a recent EE Times podcast, AI by Sally, hosted by Sally Ward-Foxton, during which she interviewed Sam Geha, Infineon’s Executive Vice President of the IoT compute and wireless business.
Geha identified the biggest challenge in the adoption process are the gaps in the ecosystem.
“What we mean by that is, starting with our end customers. Our end customers don’t have data scientists and PhDs in model development and optimization. That’s why, internally, we are trying to help them by developing what we call deep craft studio, which gives the developers the option to bring their own data and create a model or use our readymade AI models, or start from scratch by doing data collection and building their models,” Geha said. “So, for me, this is kind of what’s holding people back, and why it’s taken time. Developing that ecosystem with AI acceleration, with tools, to support the developers is the key to future success.”
He then pointed to the hardware issues and, “to the chip designers or the hardware designers, are realizing that you also need more memory at the edge in order to support these growing models and the sensor data, and in order to do that. So, to be an effective edge AI solution, the MCU must collect large amounts of data and execute sensor fusion, and have the capability to run ML models.”
Ward-Foxton then probed by asking what’s missing, and how has this been kind of holding us back? What can we do to help?
Geha used the example of a guy that’s an expert on designing coffee machines: “So, you’re the coffee maker, and what’s missing is how many people do you need to hire, how many software people do you need to hire that can say, ‘OK, in how many ways I can order coffee?’ Well, there’s maybe 500 total words where I can say latte with sugar, with cream, or whatever, and maybe in 100 languages.”
“So now the question is, I’m really a hardware guy, the guy building the machine and I know how to mix the right ratio of coffee, but I don’t have the software experts to understand words. Infineon bought a company called Imagimob, which designed an end-to-end platform for developing AI and machine learning on the edge, and we are pairing it with our hardware product, PSOC edge. So we have rebranded the name Imagimob to Deep Craft, and we’re providing the platform with tools and flexibility for developers to customize their models to their specific needs and optimize it to work with our PSOC edge platforms.”
“So, if you are that guy that’s trying to have the solution and if you don’t want to hire these software people and you don’t have the models, we can provide you ready models. We can also develop those for you; we can go to a third party to make these models for you, and then we can integrate them. That, to me, is the next frontier here of how to solve this problem at the highest level.”
Ward-Foxton highlighted “designing an AI product, if you’re making your own model, is super challenging, data set collection, curation, and Imagimob or now Deep Craft kind of provides all of those, so all of the services, right?”
Geha answered affirmatively, adding that, “We’re even training other external companies to say, OK, you don’t have to go with Imagimob – we’re trying to open up the ecosystem where there’s a lot of other design houses, other places that can actually do the models for you.”
He goes on to point out that the language models are a strategic asset and need to be secured: “At the end of the day, the biggest value or expense for these customers is actually the models themselves….in some cases, the model itself costs a few million dollars and they want to protect it. Therefore, the hardware needs to be super secure.”
Ward-Foxton pointed out that this is, “Really about educating embedded designers about AI models and how they work and what to do with their data.”
Geha noted that they want to provide “The best out-of-the-box experience.”
That means, from going to the website, figuring out what product to buy and which kit to order, understanding how to download the software, how to secure access to either ready-developed models or how to bring their own models, and then how to design it all, including using Infineon’s PSOC edge product.
“We’re trying to make that journey easier and even create a community where people can actually discuss amongst themselves how they did it there.”
Ward-Foxton then asked about where Deep Craft stands, to which Geha responded that there are, “various models that could be immediately, call it downloaded, via Deep Craft. We launched three things: Edge AI software, branded Deep Craft, where we do off-the-shelf AI models; we also have Deep Craft-ready models.”
Next year they will add the third option, Deep Craft Studio, “which is a platform for AI model development, meaning you can actually enter the data and it helps you make the AI model. I would say pairing it up with our PSOC edge is really the focus for us for the next year.”
I recommend
listening to the full podcast for more about the changes coming as a result of AIOT, but the opportunity for Data Scientists to become essential in all sorts of companies, big and small, exists.
Edited by
Erik Linask
