 Hi, my name is Nathan Eskew. I'm an associate professor here at TU Delft, focusing on AI in manufacturing, part of the aerospace engineering faculty. In this video, we're going to start off with a high-level introduction of what AI is just to get our bearings and then be able to start digging into why we could use it in manufacturing and what the applications are. So we're going to be looking at what is AI? Again, very, very short version. We're going to look at what is AI good at today because it's a very fast-evolving study and industry in practice. So we're going to be looking at that. We're going to look at where AI is not yet up to speed, especially in terms of being used in manufacturing. Then we're going to look at where AI really is unpredictable and why that's potentially a risk but also an opportunity. Then we're going to finally end with what does this mean for you? What does this mean for manufacturing professionals? So let's get started. So very short, short version. AI is designed to perform a task without explicit programming instructions. We don't tell it line by line what to do. Instead, we have a framework that allows it to learn on its own. And if we do it right, it learns very well and can solve problems for us. So we take input data and it learns through different types of algorithms, through feedback on what it's doing right and wrong, very similar to how any human would learn a new task. And then that becomes a model that we then feed new information in order for it to make decisions, solve problems, depending on what we're trying to get done. So in terms of the terminology you might have heard, there's some different phrases out there, AI, artificial intelligence. There's machine learning, there's deep learning, there's neural networks. What do all those mean in context and in relation to each other? We have artificial intelligence, which is a broad definition that we just talked about. Within that, you have machine learning. This is where there's actual input feedback adjusting of bias for a specific algorithm where it tunes itself to become better and better at a task. And within that, there's a concept called deep learning. This is what is referred to sometimes as various neural networks. And that functions very much like our brains do and is very complex and has a lot of challenges but can do some pretty amazing things. So how do you do AI? Well, you need some different things. You need to acquire data once you've identified what problem you're actually trying to solve. And depending on how much data and how clean that needs to be, it really depends on the problem you're trying to solve and the context around that. We do what's called feature engineering, where you look at different parts of the inputs that can be separated, teased out to where you're looking at the relationships between that data and seeing how we can find those patterns, those relationships and really be able to get the correct answer depending on what we're looking for. Then we take our data and in many cases, we have the answers for an initial part of that data where we know what the answer will be. We use that for our training so we can see did it get the answer right. We set a small amount aside for testing to make sure that the model is trained in a general way but not over specific to that data. Then we have passing through a particular algorithm. That's what actually makes the decision. So in terms of AI, what problems can it solve? At the broadest definition, AI can do two things very well. One, it can provide classification. So is this A or B or describe what this is? It can tell you a lot about the current state, what you're doing. So that can be natural language processing. That can be photographs. That can be looking at data and saying, what's correlated here? Is this doing what I want it to do? Or AI can predict things. In a sense, it can see ahead into the future. It can use past and current data but it can also look at subtle different behaviors and correlations and relationships with that data and then it can provide predictions that can, depending on the model, be extremely accurate and helpful. But you have to be able to set this up with the right kind and the right amount of data depending on what problem you're trying to solve. So what is AI good at today? Well, there's object detection. It's getting better and better at saying, hey, an object's here. So looking at things like navigation or mapping things out. It's very good at anomaly detection and this is especially important in manufacturing where through different types of sensor but especially just visual data through video and photographs, it's able to see, hey, this doesn't look right. I think there's a problem. Then there's also process and safety monitoring and prediction so it can look and say, hey, this person isn't wearing their safety gear. It can say this situation isn't correct or hey, I think there's a risk that something can go wrong. So a quick example of this of what AI can do, especially when you put together different types of AI to solve a bigger, more complex problem. So let's look at supply chain. Say we have a large organization, in this case, in the U.S. and we have many different suppliers all over the place and it's very, very complex and difficult. How do we keep track of it? How do we make sure that we are taking the right action at any given moment? Well, we can take a lot of that data and we can take past data and we can provide supervised regression. So we know, hey, there's very high probability that something is going to happen. We need to take action if we don't want that to happen. Then you might have what's called an expert system where we have different types of data put together with different decision trees or if-then type rules where we can see, hey, based off of this information, this is probably what's going on. And here's the core data that tells me why that is and here's the probable action that I need to take. Then you might have documents that feed in that data that can be scanned in a more loose way to be able to find correlations, pull out different names of people, different locations, and that can help track down maybe who do I need to talk to. So if I see that there's a problem, I identify the root cause of that problem, I need to know then what action I need to take. All of these things can help get me to the person I need to talk to very, very quickly versus days or weeks or even months. So where does AI fall short? Well, in a lot of cases, AI is statistically impressive but individually unreliable. And sometimes that's fine because we need a mostly good solution, but sometimes it's not fine unless we have as close as possible to 100%, especially things like safety, precision manufacturing. So we see that also with AI, the cost of failure is measured in a way that even if we know statistically the AI is better than a human doing the job, it could have worse implications if the AI fails, even if it fails less than a human does. So where is AI unpredictable? There are different case studies that have been done, but AI has been able to find ways to solve the problem that are not obvious at all to humans. There was an AI hide-and-seek game that was put together by different bots. And what they found, and it's a fascinating read, what they found was the AI bots learned to break the laws of physics in the game. They learned that if they hit a certain corner of the arena just right, they could fly into the air and fly over walls. Another found that it could actually climb on a box and break the physics engine and move around on the box in order to accomplish the goal. So that's scary if you need it to work the way that you expect, but also very exciting if you want the AI to discover new and innovative techniques. So what does this mean for manufacturing? AI is already being successfully deployed in manufacturing in many different ways. It can do amazing things, but it all comes down to knowing what problem you're solving, making sure you have the right data, feeding it through, training a model, and then deploying it correctly. It's not ready to perform all tasks, at least yet. At some point, most other problems will be solvable with AI. It's just a matter of time and the amount of effort that we are willing to put into that. It'll always be somewhat unpredictable. In a sense, so are we. So we have to keep that in mind and prepare accordingly to make sure that we really are solving the right problems with the right risk profiles in mind. If you have any questions, please use the course chat discussion, and I'll see you in the next video.