What's the difference between AI and Machine Learning? - With James Loach

Trend Detection Podcast ·

Show notes

Welcome to the Trend Detection podcast, brought to you by Senseye Predictive Maintenance – which gives you visibility and insights into all your assets, from single machines to full plants to help you reduce downtime, increase knowledge sharing and accelerate digital transformation across your organization.

You can find out more about how Senseye Predictive Maintenance can reduce unplanned downtime and contribute towards improved sustainability within your manufacturing plants, by visiting: www.siemens.com/senseye-predictive-maintenance

Transcript

Welcome to the trend detection podcast, powered by Siemens, and now your host, Liz McGinn. Hello, everyone. This is Liz McGinn with Trend Detection podcast. I'm here joined by James. We are going to talk about machine learning and AI and what's the difference, and he is the teacher today. So, James, you want to kind of introduce yourself and why you're qualified to speak on the topic? OK. OK. Hi, Liz. Yeah, so, um, I'm James Loach. I'm head of research at Sensai and have been since the early days of the company, um, and I'm also, um, part of the, the product leadership team. You know, as to why I'm qualified to talk on, on this topic, I think this would be, um, something down to practical experience since I've been building machine learning AI systems, you know, for, for, for that period. Uh, my background before that is in something different though, which is experimental particle physics. Um, but in terms of the way you think about the world and the actual things that you're doing, it's not necessarily too, too different from what I'm doing now in the machine learning and AI, you know. Nice. That's interesting. Um, I guess to jump into it, when we're talking about artificial intelligence and machine learning, can you give some definitions just so everyone's on the same page when we first start? Yeah, I think you can think of artificial intelligence as a thing that you're, you're aiming for, right? So we all know what intelligence is, you know, for our day to day interactions with intelligent beings or animals or all this kind of stuff, right? So this is the kind of thing that that you want. So a system which is able to take in sensory information from the world and based on that information take intelligent actions in the world and, you know, continue on like that. And machine learning is one way to, um, achieve that. Yeah. And machine learning, um, it's about having systems, um, that, yeah, learn, learn from data they take in from the world, right? So it's, it's a model that you're training on data from the world and this model will make, make useful predictions. And, in the case of something like an animal, um, in a natural system, these will be the actions to take. Um, and in artificial systems, you know, the output of the machine learning system could be, um, a forecast, it could be a suggestion of, um, what, what the problem is with your machine or in something like chat GPT. It could be the, the message from the, the helpful assistant. OK, so AI is the goal, and machine learning is one technique that, that we use to, to try and realize that goal. OK, so machine learning is used to build artificial intelligence. Yes, yeah, awesome. Um, I guess That kind of skips my next question. Um, do you, so I noticed that a lot of these terms are used interchangeably, um, with artificial intelligence and machine learning. Um, is that, um, correct in the assumption? I know, like, technically it's building up, but like, when we're talking about it in the media or within, you know, products or anything like that, um. How do you go about it? Yeah, I think at this, this point in history, the terms are almost synonymous, right? So they are formally different things, but you do not come across many systems in the world now either claiming to be AI or trying to achieve that that are not driven by machine learning. OK. This is a little bit different historically. So if you go back to the 60s and 70s, you know, particularly, there were, there were other, other approaches. Yeah, um, you know. The, the traditional other kind of approach is where you, you're sort of programming in the rules, right? So you can imagine such a system, you know, if the user says this, you process it in some way and then you give an answer and you, you have like complicated rules or scripts or things like that. And yeah, the thing you do with machine learning is to get away from all of that, so you don't program that stuff in. You have the system learn what the appropriate responses would be, yeah, but these days we can make a subtle distinction between AI and ML, but you know it's not like a huge practical importance for the reasons I described, yeah. Nice. Can you give a little bit more of the history and the journey, how we got to AI from machine learning? Um, OK, well, I mean, the, the, the history of AI is, is difficult to sum up very briefly, but, you know, I think the important things were I can pull out free from, from what I just mentioned, right, that people are starting out with, um, You know, machines which, um, work in a symbolic way, right? So, you know, you have different concepts and operations on those concepts and you kind of, you know, get, use, use information from the world to like, to program all this up, right? You know, so you can describe these objects in the world and they have kind of rules about what they are and how they interact and things like that. Um, and you can build up. Yeah, systems which feel like they've got a certain amount of intelligence in a sort of deterministic symbolic way. And the thing that I, I, I, I mentioned in my previous answer was, um, yeah, famous system, I think from the 1960s called ELISA, yeah, which is like. Like a therapist, you know, that you can, it will ask you how you are and you can write something and it can infer from that like, oh, tell me more about that, uh, tell me about your family, or what do you think about that, etc. And at the time, people could be sort of convinced that there was, there was some intelligence behind this, but all rule-based, you know. Um, and what we've seen from that time to now is this other approach. You have to start off as something which is conceptual, something which works in toy situations, something which has had various problems in making it work at a larger scale solved, you know. And you have, you have this sort of theoretical progress over a number of decades and at the same time, you know, very interesting changes in, uh, com, um, hardware, right, that, that make it convenient and, and possible and possible. There's a whole story there, right? But, you know, you, you have this, this movement from early success with symbolic rules-based systems to the sort of absolute dominance now of, of machine learning things, um. Yeah. Nice. So very input. If it says this, then you respond with this rule-based versus now it'll almost think on its own. Um, yeah, well, it's using, it has, it has, in a weird sense rules inside it. I mean, like an AI system is operating in a deterministic way, you know, it's this thing with all of these weights and connections and you push your question through it in simple terms and you get your, your answer out. So there's something deterministic in that. But the key point is that human beings didn't write it, you know, and there's a whole industry now of trying to understand what is going on inside, you know, cutting edge AI models. People have to study them in the way that you might study your brain or something like that and try and figure out how it's actually doing what it is doing. Um, yeah, anyway. Interesting. That's interesting how we went from, um, kind of that role base to now we're trying to figure out how it's thinking on its own. Yeah And, and, and this is quite difficult, yeah, but, but it's, it's somehow easier than working at how brains work because of course you can, um, you can do brain surgery on AI models very easily, yeah, and mess around with things and, and see what happens. That makes sense. I mean, brains are very complicated and it's hard to study your own thing when you're a brain yourself. Awesome. So when we need to translate that to predictive maintenance, what does that look like? Um, OK, so. Yeah, in, in, in, in, in broad terms, you know, you've, you've, you've got some machines, right? Um, they have some history themselves, you know, how, how they were constructed, you know, and that they're doing a particular task. They have some particular history in your factory, you know, because people have done repairs on them, you know, they've been used in particular ways over periods of, of time. Yeah. So you have all this, this historical stuff to do with your machines, and then you've got a particular thing at the moment, right, which is something is happening on your machine. And that could be something you're observing or it's something in, in the data that you're recording. But, you know, you've got this whole, whole set of information and what, what you're interested there is, um, is there anything happening here that I should be worried about? OK? And if you, that's true or you just are worried about it, what, what should I do? OK. So you have this whole load of source information basically, and what you want at the end is something like a, a statement of whether this deserves your attention and if it does, what you should do about it. OK, um, and so you've got, you've got the box in the middle that's taking that input and producing that, that, that output. Um, I guess we were going to this a little bit, but Sensii is, um, a whole complex arrangement of different things in that box. But you know, in. Uh, simple terms, again, in, in an old-style system, you could imagine that box being full of rules, right? And I think a lot of maintenance stuff was, was really done, done like this, you know, if the, the, the value of the current crosses a a certain value, then it probably means this or this sort of thing, right? Um, and then on the other extreme, you can imagine having that as a pure machine learning system. Again, so maybe you, in the limits, you don't know what's going on inside it, but it just reliably and usefully maps that contextual information onto statements or actions, uh. Advice, you know. Nice. That's interesting to hear. I mean, it's um. Definitely a lot of people go about the rules in predictive maintenance, not quite predictive, I guess, but prescript, not prescriptive, sorry, in the way that they are looking at it and they don't want it to go over these thresholds. Um, so when we're using, say, machine learning or AI, how does that train, like, I know you have this box that's giving you different rules, but why, um, is that approach? Different or better, I guess. I don't want to use the term better, but I'm kind of thinking better in my head. Um, well, it, it, it, it's just very complicated, um, answer that question because you, you cannot at, at present, you know, certainly for, um, the sort of very, um, machine agnostic application like Sensi, you cannot train a machine learning model that can just convert that arbitrary set of input information into, into, into reliable output, you know, and do so quickly and, and at scale. Right? So what, what you do inside that system, um. Yeah, I'm not sure how deeply to go into sense I hear, right, but you, um, yeah, so we have, we have a complex arrangement of things, um. Maybe we should start this one again. What, what's, I suppose, what, what specifically do you want me to? You're good. I was just curious, I'm kind of leaning into the differences and kind of the benefit of using AI or machine learning, uh, in comparison to kind of that traditional rule-based following. Yeah, OK, OK, OK, OK. OK. So when using AI and machine learning to for predictive maintenance, how does that, um, Change the the viewpoint of kind of this rule-based following, so the form of, um, you know, kind of that black box being filled of questions and answers versus machine learning. OK, so, so I think you think this is to do with optimal use of information. OK. Um, so, you know, if, if you've got some rules, I mean, somebody has to have, have come up with these and they, they might be sort of sensible rules, but they, they are, they are quite crude things, you know, like if it's above a certain value, it means this. If it's below, um, it, it, it's not a problem, right? But you know you can imagine the data has all kinds of values, and you may see the initial indications of some problem developing before it crosses the line. And then even if it's crossed that line, you may see in the history of the asset that actually if it's only just a little bit above it, it doesn't matter. But alternatively, in another situation, it was just a little bit above, that really is a really big problem and what have you. So, I mean, it's just a question, you have a lot of complexity and subtlety in the real world and you are representing that in a rules-based system as simple rules that a human can understand, but there, there are abstractions and simplifications of the world. And I guess the, the promise of machine learning would, would be that you can by learning, um, based on all the, all the things that have happened and all that you know. And that you can just do this in a much more subtle way, right? Picking up on all the little, um, little changes in the data, all the historical things and just very optimally say, OK, I think now is the day you need to go and have a look at that for, for these reasons. Yeah, so you, you do, you do that more in a, in a more optimized way. You make better use of the information, and this means you make better use of the user's time and, and waste their timeless. Yeah, that would be the promise, right? But in, in reality, as I said, you, it's difficult at the moment to build this pure, um, Machine learning system to do that, but everything is moving in that direction in a sense like our system is a complex mix of AIML things and statistical things, right? Mhm. Yeah, plus human input is important for, I guess, the machine learning as well for our systems. Yeah, and that's super important and again something something we go on to, but you know the system really, really wants, wants, wants to know about the, about the world, right, how it was in the past and how it, how it is now and doesn't have direct access to that, you know, you have access through things that users write or time series that you collect or maybe a photo that the user takes or something like that. But at the moment, and this may be different in future, the user is required, right? It's. The entity that really tells you what these things in the data mean, right, so you can learn what they mean and, you know, to do this task in the future, yeah, yeah, because it's always learning. Yeah, yeah, yeah, but it has to get, it has to get ground truth knowledge from somewhere and Senii is, is. Works around the user because the user normally knows, right? They're in the factory, they've been working on this machine for 3 years or what have you. They have some tacit knowledge. You, you want to get that information out their head and get it into the system and then use that knowledge, you know, while the user is doing something else. Yeah, so do you think um AI will ever kind of replace that human or do you think a human will always be needed for this kind of thing? Um, it, well, if, if you, you say always, I think, I think the answer is, is no. Um, I mean, it, it. Because it depends on your worldview at, at, at some level, but I, I would see human beings as fundamentally machine-like, OK? And I don't see any particular reason why anything that human beings do can't be, can't be replicated artificially, um, and then ultimately done a lot better. OK? Um, and then, you know, if, if you're in that situation where you can very cheaply. Um, and accurately to do all of these kind of things with a robot. Um, at that point, humans would be doing it as a hobby, I would think. OK, right? Um, much like we, we continue to play chess even though, uh, you know, chess computers can, can be even the best chess players. But anyway, but, but that, that could be a very, very long way, way away. Yeah. And where you have, you know, expensive, imperfect AI systems, obviously good reasons to keep, keep human beings involved. And even though you can kind of see a fuzzy route to complete a replacement of Um, manual labor and knowledge work like our jobs also, right? You can kind of see a fuzzy route to that, you know, I think breakthroughs would be needed and the timescale for this to happen is still quite, quite uncertain and could be long, yeah. Well, I'm glad it's got a minute. I have to reach retirement before they start replacing my job, hopefully. Um So do you think that um when you go back to the, the manufacturing, like, vision of it, I know that it's kind of hard to talk to um people like are resistant to change, similar to like kind of manufacturers are kind of usually slower to, uh, you know, move towards the direction, but I'm seeing a lot of changes in that, um, not like, quickly, but like they are kind of leaning into this AI transformation. Why do you think um this one's being Like Um, except it's easier, or is it not? Am I just, well, I, I think people are smart, right? You know, and when, I mean, the, the current AI systems based on language models are really very useful. And, uh, yeah, I mean, I, I use them all day every day. Uh, I expect most people listening to this probably to also, or at least a, a good chunk of time. Um, when you have something that is real like that and so obviously useful. Um, yeah, it's, it's no surprise people are interested adopt in adopting it and working how you take that sort of raw power and intelligence and somehow get it to work, you know, properly in your particular environment, if that's your, your factory or something like that. OK? Um. You might contrast that to before language models where machine learning was still used in factories for all kinds of tasks, but this is more narrow AI, so doing like very specific things and you have to go through, you know, maybe long expensive training processes and keep rebuilding, you know, and have data scientists involved all the time, this kind of thing. Yeah, so even though there was something there that, that, that was working and you could show off demos, yeah, very expensive in general, and the reality of those systems is often something of a disappointment, whereas most people's interaction with language model-based eye is kind of. It's wearing off a bit now, but some kind of awe, it's like magic, yeah, and so it's, it's a very different headspace, yeah, I think. Yeah. So what did it look like before we went into kind of this AI, like the more basic version of what you would see like in manufacturing for, um, those kind of questions and answers? What were they using it for in there? Yeah, I mean, I suppose you could imagine things like spotting defects, something like that. So you maybe have some vision system and you have lots of examples. This is a good thing and this is a bad thing, you know, different categories of bad thing, and you know, you collect all of these images and you have to then, you have to train a model and you've got to commission it and then you have the problems if the lighting slightly changes in your factory, and you've got to be resistant to that and you're doing all that kind of thing, yeah. And then, yeah, the, the reason it's more exciting with systems based on language models is, yeah, like you log into such a model and take a photo of your product and ask it whether it looks OK, and the system immediately has some opinion on that, which is quite robust. You know, it doesn't replace that previous system, but, but there's the sort of power there that you get, get for free, right? It feels like it's for free, and then it's like trying to, to tame this beast, you know, whereas to me, like traditional machine learning systems is like, you know, climbing big mountains is very hard, you know, um, yeah, to, to, yeah, to make these things work and be robust and what have you, like a hard grind. You don't feel you're getting free magic from the world somehow. Yeah, you have to keep correcting it pretty much anytime something's a little bit different. Or making it robust, yeah, all that kind of stuff. …

Full transcript (WebVTT)