Issue link: https://iconnect007.uberflip.com/i/1502623
18 SMT007 MAGAZINE I JULY 2023 is quite remarkable. It's that element of sophis- tication when I refer to AI. Johnson: There is also AI doing pattern matching: looking at a huge pile of data and sorting out patterns for a purpose, whether it's diagnostic or quality inspection. That's not generative, right? What do we call that kind of AI application? Typically, we call that machine learning. Say you want to program a machine to look at a mil- lion images of something and learn to do a task based on its analysis of those million images. I might not understand how it accomplished the task, although I would understand the answers it's giving me. It makes sense because you've analyzed so much data. at's machine learn- ing, sometimes called deep learning, which is when neural networks carry out the task. Deep learning is a subsection of machine learning. Generative AI is an element of deep learning. Johnson: Specific to DarwinAI, is this AI or is it machine learning? We fit more in the machine learning/deep learning realm of things. We are leveraging our latest tools in deep learning to do visual inspection of PCBs in a compelling way. For most EMS companies, there isn't a lot of data on "bad" things: tombstoning, polarity mis- matches, missing components; you have to collect the data and it's painstaking. So, we are leveraging generative AI to generate syn- thetic data across a broad range of PCB config- urations to make our AI more effective. We're now doing both and that's been a recent devel- opment for us on the generative AI side. Johnson: As one refines the scope of the AI application down to some very specific and closely defined applications, such as AOI, does that make AI development easier? It certainly makes it easier than the devel- opment of a general system. ChatGPT is a multi-million-dollar effort to train a gargan- tuan model—it has 175 billion parameters. e technical process of creating an AI specifically for a PCB with all its nuances is a more tracta- ble problem than creating ChatGPT. But we sometimes forget that AI is only 30–40% of the system you're bringing to mar- ket. ere's the hardware element—the cap- tured images must be high enough quality so the AI can make sense of it. at is an effort distinct from the AI. e AI can be great, but if the images are blurry, they're not high resolu- tion, or it doesn't work with their workflow, it won't be an effective product. Andy Shaughnessy: What's the biggest advantage of AI for your company or your customers? It's a couple of things. Its accuracy at detect- ing anomalies and irregularities that would be false flags for many traditional AOIs. e other big advantage is how quickly you can config- ure a new product. Give us a golden board and within one or two minutes we have a full under- standing of the board. So, it's how quickly the AI can understand the new product configura- tion and get you up and running. A traditional computer vision system requires explicit programming, whereas AI gets smarter and more effective over time. To do a pixel-by- Sheldon Fernandez