Issue link: https://iconnect007.uberflip.com/i/1534953
94 SMT007 MAGAZINE I MAY 2025 edge base. In our use case, the input will be images, and the output is used to determine whether this image shows a good or defec- tive solder joint. There's a lot of discussion about AI language models in a particular application. Are you specifically looking at image models? Not necessarily. I am focusing on federated learning, a collaborative learning approach where multiple companies collaborate. It is applicable to every type of machine learn- ing. It's potentially usable for classification on numerical data and on image data. It's also possible to do language models, of course. It seems that a critical challenge is getting good data into your model in whatever way that might be accomplished. Are you using proprietary or customer data from multiple companies? I work at a university, and this is my research project. One of the project partners is the manufacturer of inspection machines. e data originates from multiple customers of this company. ey all have the same machine. at's why this is easy to implement; the data is always structured similarly. is is, of course, also applicable to more complex use cases. When we have different types of data from different manufacturers of different types of machines, then it gets more complicated. In our case, data quality is different between the companies. We definitely see a significant effect from that, but the structure of the data is the same. To create a good AI-driven inspection process for assembly, having access to lots of data from multiple sources, multiple compa- nies, and even multiple approaches makes it robust enough to predict well. Pulling from just one company may not be enough. How do you see the handling of proprietary data? In this approach, we use data from multiple companies. Also, we create models that are more robust and more generalizable regard- ing unseen types of data. e whole concept of this federated learning approach is that data always stays local, where it's created. Every participant trains one local model themselves. en, instead of sending the data to a cloud server, we send the models. From these models, it is basically impossible to reconstruct data. at's why data is always protected, but we can still use the knowledge from these information sources. That's pretty powerful. Where do you see applications for this approach? How will it change the EMS industry? It can be applied to many of these cases across every type of industry. It's always a good idea to consider this when higher quantities of training data are desired. When problems are more difficult, it's usually the case that these deep learning models perform better with more data. e second requirement is that we want to keep our data private. is is true for most of the use cases in production. Manufactur- ers want to keep their data private, and that's why this approach is generally applicable for a lot of different use cases. One of the key themes in electronics manu- facturing is the need to collaborate and pro- tect intellectual property. This application is spot on. is approach is not new; we didn't invent this approach. Currently, it's applied to things like when consumer electronic devices interact with a server, for example.