Building an AI foundation for a smarter digital enterprise Podcast – Transcript
Bringing AI into industry isn’t something that can happen all at once, rather, something that will happen gradually by applying AI to individual areas where it can have the most impact. With that said, as these foundations continue to grow, more complex, overarching AI applications will begin to find their way into industry as well, offering greater flexibility and possibility then traditional systems can.
Check out the full episode here or keep reading for a transcript of that conversation.
Spencer Acain: Hello and welcome to the AI Spectrum podcast. I’m your host, Spencer Acain. In this series, we explore a wide range of AI topics from all across Siemens and how they are applied to different technologies. Today, I am joined by Dr. James Loach, head of research at Senseye Predictive Maintenance. In a previous episode, you mentioned that time series foundation models are already in the hands of customers with great success and have been for the last few months even. So where do we go from here? What’s the next step for you at Senseye? Do you consider time series foundation models for error detection and anomaly detection to be a solved problem? Is there like a place to go from here or what?
James Loach: Yeah, I mean, on a narrow question of forecasting time series, I think we consider it to be solved. Like, basically happy with every forecast it generates, the only thing that we might be interested in doing there is kind of reducing the cost. You know, so we will be on the lookout for lighter model architectures in the future. But Senseye is basically a two-step system. It’s got a part of the system which is… you know, doing analysis of streaming time series data. Yeah, looking for interesting things, unusual things, things that matched stuff in the past, you know, looking for data that implies a future threshold for violation, all this kind of stuff, right? So you have a huge amount of data coming in and then as part of the system is trying to look for things of potential user interest, okay? And that stuff is fed through to a second part of the system. You know, where essentially simplifying a lot, the results of that are fed into language models, okay, you know, with all sorts of other information, to try and really decide whether this is something that should be raised to the user as an alert, and when we raise it as an alert, you know, we want to give a very clear description of what happens and all the possible causes and things like that, you know. So, originally, the first part was all statistical and the second part was like very rules-based. Okay, and we evolved that to a situation where the first part was mostly statistical and the second part was using language models. And I think we talked about this in the past, but this is, they proved very, very useful to us in that second part in really making our idea of the system work as we wanted. And yeah. You know, and the progression you’re saying with these kinds of things we’re talking about is bringing more foundation models and AI into those initial parts as well, right? So at the moment, there’s a mixture of statistical models and this time series foundation model chronicle we’ve been talking about. And I can imagine using, yeah, it’s time series foundation models for more activities in that initial part. Okay. But again, it’s a lot of data, it’s streaming data. You know, we need to make quick, light decisions about things. And with foundation models, that was always going to be the challenge. So we’ve put it in the place where it makes the most economic sense at the moment. And yeah, we will be trying to push this into, into those other areas. Yeah, as the technology allows, yeah.
Spencer Acain: Well, it sounds like there’s all, as always, there’s a lot of room to grow. So kind of to take a step back from this, it’s these time series foundation models are many ways a solved problem in your domain here, but what if we take a look at this in kind of a broader context? You know, you talked earlier about the way that it’s that these transform models are looking at data through embeddings and how they’re able to look at things that might at first glance not seem all that similar and say this is this is like this and then which send that to the users. So how do you see that expanding? You know, these LM methods at large kind of you know helping across enterprises, you know, like with larger data sets maybe beyond just with the type that Senseye is working with?
James Loach: Yeah, I mean, that could be a very big topic. And the narrow topic of embeddings and time series, you could think of this as a rag for time series. RAG is this retrieval augmented generation, where you have lots of your bits of your documents or text, they’re embedded. You have a system which does matching across that and can use that to help you answer questions. You know, so, you know, HR system, you ask something about some sort of policy or what have you, right? And, you know, it can embed your question, can look through the whole database of HR material, right? Looking for a piece of text which has an embedding which is similar to your question, make it simple terms and then pull those in and then kind of summarize those back to you to sort of answer your question. So that’s a very common language model paradigm, right? Rack, yeah. And yeah, for the time series, you can use embeddings in this kind of analogous way, right? So you’re using it to index time series and retrieve time series, right? And so you’ve got some time series and then you have all these previous time series with some kind of label, right? A thing happened or some work was done or some problem or whatever it be and you’re just using it to be able to retrieve incidents from the past that look similar, you know? So there’s sort of a narrow use of embeddings, but you know, within somewhere like Siemens, yeah, you know, like Time Series is one of the fundamental ways that we end up getting information on what’s happening in the real world. You know, you have sensors and they produce time series and Siemens has this colossal amount of such time series data. And a lot of understanding about what that data signifies and what it means and those consequences are and stuff like that you know um so this is something yeah if you’re dealing with hardware and time series and you want to tell what’s happening or predict what’s happening. This kind of rag time series thing is sort of makes, makes complete sense. But you know, if you’re talking about like foundation models and embeddings, how they could be used or are being used in industries or business, this is just like a huge sort of giant topic. But yeah, it is an effective way of getting information from huge databases and huge collections of information into the context of language models for them to reason over and to summarize for you. Yeah, super, super powerful. It is everywhere and it will be everywhere. Yeah, they can encompass different modalities as well. Then if that answers your question.
Spencer Acain: Yeah, no, thanks. That’s a good answer, I think. Thanks, a lot of insights there. So I guess as we kind of start to wind this down, you know, we’ve talked a lot about, we talked a little bit about foundation models, talked obviously a lot about Sun-size time series foundation model, but how does this compare with the broader Siemens efforts? Because of course, you know, we are talking, you know, broadly about the Siemens Industrial Foundation model. Is that an analogous, uh, you know, a piece of technology is that, would that encompass the same as what you’ve done? Or is this, or are they kind of working towards separate goals here?
James Loach: Yeah, I think you can think of the general idea of industrial foundation models as being some way of a broad umbrella term, you know, for many different systems and interacting systems of which the thing that we just described to this Chronicle model is just sort of one example of one class of thing. And yeah, I mean, the way you could make the connection here, I mean, Chronicle is taking a base model, you know, that’s trained on data from the public domain and the open internet and things like that. And it’s specializing the training for industrial data, yeah, that’s not widely available on the open internet. And then is restructuring the model in ways that, you know, enable it to solve an industrial use case. And I think there are many opportunities for doing something similar to that, as well as other things, yeah, of specializing open, trained or openly available base models on this huge kind of corpus of knowledge and information and expertise, time series, images, all these kinds of things that Siemens has. And you know to do with it specific to domain right and Siemens will have a lot of information yeah which is not available on the open internet and not available to hypers who are training chat to chat pt and things like that yeah and so there’s like a I would see it a sort of a specialization thread to it And then also you can imagine training all kinds of interesting classes of model from scratch internally for different kinds of, using different kinds of industrial data and for specific kinds of industrial tasks. But yeah, an industrial foundation model to me, it’s an umbrella term, yeah, to do with specialization of open things and internal construction of things, yeah, in this sort of domain-specific way. And yeah, Chronicle is, you know, one, you know, limited interesting little example of that, yeah.
Spencer Acain: Yeah, all right. So it sounds like, in a sense, we’ve already contributed to the industrial foundation model with this, creating something that works well in this set of industrial data. So to kind of round things out here, I was going to ask, you know, what are your, any closing thoughts on this, you know, either about what we’ve just been talking about or where do you see the broader role of AI going in industry or in maintenance and predictive maintenance or data forecasting in general?
James Loach: Yeah, I mean, I think it obviously has an extremely central role. And you know, when I described a little potted history of the Senseye app earlier on, you know, that is a story of foundation models and modern AI. You know, eating larger and larger bits of our application, basically, right? And that’s something that I can imagine continuing. Yeah, the models becoming more powerful, faster, cheaper, able to compass more modalities, right, and you can push more and more reasoning, you know, into these central foundation models, you know, and ultimately remove a lot of the complexity that you have in traditional systems, which, you know, a good bulk of, you know, the underpinnings of sense size in that sort of traditional layer right um yeah so um I think yeah that that’s sort of a pattern that that that will continue um yeah and like um I it’s probably not worth me saying too much about like the wider question I mean you know these models in general are extremely powerful, underlying technology continuing to increase in power and people are still figuring out the best ways to integrate these into products. Chronicle is very nice because it’s taking this particular form of technology. And it’s, you know, it’s very aggressively turning it to a practical problem and getting it in front of customers, doing something useful. We can learn what the customers like and what they don’t like, continue to iterate and things. But yeah, this is just one example, obviously, of many such processes going on, you know, throughout Siemens and, yeah, industry more broadly. But yeah, it’s just a very interesting time and, yeah.
Spencer Acain: Yeah, we’re really on the cusp of seeing what this technology can do and where it’s all going to be applied. It’s just almost too early to say in all the ways it’s going to be impactful.
James Loach: Yeah, it will take quite a while to play out, I think.
Spencer Acain: Yeah. Oh, definitely. But I think thanks for that great response. I think that’s about all the time we have for today. So once again, I’ve been your host, Spencer Acain, on the AI Spectrum podcast. James, thank you for joining me today.
James Loach: Yeah, thank you, Spencer.
Spencer Acain: Tune it again next time as we continue learning more about the exciting world of AI.
Siemens Digital Industries Software helps organizations of all sizes digitally transform using software, hardware and services from the Siemens Xcelerator business platform. Siemens’ software and the comprehensive digital twin enable companies to optimize their design, engineering and manufacturing processes to turn today’s ideas into the sustainable products of the future. From chips to entire systems, from product to process, across all industries. Siemens Digital Industries Software – Accelerating transformation.
