On this episode of Heads in the Cloud, James and Joe talk AI and machine learning technology and trends.

‍

James Burnand:

Hi, everybody, and welcome to this week's Heads in the Cloud podcast. This week we're going to do something a little bit different. We had a guided topic that we were going to cover, and Joe and I have both unanimously decided to not do that topic today. So we're sorry, GI, who's probably watching this wondering what the heck we're doing. We're sorry, but we promise we'll do that one next week. For this week, what we really wanted to talk about with some of the industry trends, things that are happening, and specifically, Joe and I, we're going to cover off some background and some information about AI and ML. So Joe, if you want to maybe kick things off, we can just talk in generalities about what that is.

Joseph Dolivo:

So starting out with a definition, artificial intelligence and machine learning, so machine learning is a subset of artificial intelligence, but they're topics that have been in the news for a long time. And it's everything from using vision for self-driving cars to things like ChatGPT, or Stable Diffusion, and OpenAI, and DALL·E 2, which are super fun to play with, by the way. So it's interesting to see how quickly things have evolved. And one of the things that I like about the evolution, especially of things like Stable Diffusion in DALL·E, so the image generation, is that a lot of that growth was driven by some of these tools being open sourced.

And so all of a sudden you take something that was previously done by proprietary research groups inside of a company, they're open sourced, and then the public at large is able to take those and start building incredible things with it at an incredible rate. And it is industry transforming, and I don't think it's an exaggeration to say that, although there are some exaggerations with certain things about industries going under and jobs disappearing that I think are not quite the case, but it's going to be an interesting few years, to say the least, I would say.

James Burnand:

And I don't know that everyone realizes just how pervasive AI has become inside of your daily lives, everything from auto completing your texts, to predicting what sort of traffic you're going to have when you type in a route into your MAP software. The processing capabilities of our cell phones and of our computers and the devices that we use have been in many cases optimized for running these workloads, but also there's been a lot of pragmatic and practical applications in our daily lives that we probably now are starting to take for granted.

But certainly if you look at the grand scheme of life and where these technologies apply, it's probably underserved in the manufacturing side, and there's a whole lot of companies, a whole lot of energy and effort trying to figure out exactly what the holy grail is for manufacturing with AI, including deep learning. So where machine learning is a subset of AI, deep learning is also a subset of AI, where generally unstructured data sets to create and find the insights rather than an ML where you have some idea of what you're trying to accomplish and really trying to optimize that decision-making. Joe, what do you think about AI and manufacturing? Where have you seen it work? Where have you seen it been a challenge?

Joseph Dolivo:

I think the canonical examples that people talk about are doing things like predictive maintenance, that's AI 101. And a lot of the case studies we'll see from some of the vendors, even companies like AWS who have some products that are targeted at that space, will focus on that, and you'll hear it called anomaly detection, which is a little bit of a more abstract concept, but looking for something that is out of place. So looking for defects in vision applications is a very typical one. Whereas, if you were doing traditional vision, there's a lot of things that traditional vision systems are very, very good at, finding things that haven't exactly been seen before like defects that might not be known, those are actually really good use cases for some of the deep learning tools like ViDi, which was a startup that Cognex bought a few years ago now. So that's a pretty common one.

Also for forecasting, I remember doing a talk on this probably eight years ago about a golden batch system when you're trying to predict certain batch variables or process variables as a batch is going through based on looking at past batches. So doing prediction of how a process variable is changing over time based on how it's changed in the past, again, that's a somewhat straightforward to talk about, but very practical use case for machine learning within the manufacturing space specifically. And then in our daily lives, we tend to see it with recommendation engines. So you're watching a show on Netflix and then you see recommendations for other shows based on what you've liked, or what you've watched, or what folks that are watching similar shows as you or that are in similar demographics as you have been watching and make recommendations based on that. So I don't know that I've seen a strong recommendation engine example in manufacturing just yet, but I'm sure someone will figure out how to do that soon.

James Burnand:

I hope all of our viewers that are watching this aren't distracted by the fact that the giraffe is moving the entire time Joe is talking, because Joe's a handsy talker sometimes, and the giraffe keeps coming in and out of frame. And I apologize, Joe, I was definitely listening, but there's a lot of giraffe movement behind you.

Joseph Dolivo:

I think it makes it more lifelike. So he's got some strong opinions on this, and it's actually my chair that is bumping into him, so not my hands, although, I do this, but I'm trying to not invade the giraffe's personal space too much, so he's just bouncing around.

James Burnand:

That's very respectful of you. And you talk about recommendation engines, and obviously, again, it's something we see pretty commonly in our daily lives, but also voice recognition, and specifically those of us that have tinfoil hats that we like to wear around and believe that everyone is listening to us and there's some sort of big conspiracy behind everything. I would say, I'm not quite tinfoil hat territory, but I definitely know that my phone is listening to me. And based on some of the recommended ads I get and the conversations I have several hours before that, again, this is one of those kind of use cases that has hit the personal world. But I think one of the use cases that I haven't really seen yet, and maybe our listeners have is things like being able to specifically in process control to have better dynamic adjustment of batching.

I know that we had done some work years ago using statistics on optimizing batches and being able to intervene with formulation changes in flight in the batch to try to optimize the outputs, but I haven't seen a whole lot of tools out there that have really gone to the effort of optimizing those batches and really looking at getting away from what I'll call standard and strict formulation into more dynamic formulation within some sort of boundaries and with some sort of quality parameters at the end that they have to meet. To me, I feel like that's a really huge use case that just I haven't seen the practical application of that in a lot of places.

Joseph Dolivo:

I think the reality is so many companies that are out there, they don't even have connectivity to their data sources in a lot of cases, and so machine learning ends up being a really, really series of step changes. To be honest, outside of getting access to the data, which is of course step one, there's a lot of companies that I think would be very well served by very simple statistic models. So statistical models, simple math, that can get them most of the way there. You don't always necessarily need to embrace machine learning, if you will, although there are very good use cases for it. And to your point, James, I think we're seeing some of it. I think there's a lot more opportunity for it as well. Things like ChatGPT, while fascinating, I don't quite know how that would be applicable yet, but maybe somebody's going to probably come up with something.

James Burnand:

I'm pretty sure I saw a post that someone was trying to get it to write ladder logic, or maybe that was CoPilot. But conceptually, if you go back to how you go about deploying and managing industrial systems, there is an argument to be made that there's a lot of repetition and a lot of standardization that could be done with the way we design systems too. So there may be maybe a use case there that maybe hasn't become pragmatic yet, but will be.

Joseph Dolivo:

It is even an interesting tool for helping with programming. I think just like anything coming from, let's say, I'll pick on junior programmers, you want to basically do quality checks and make sure that it's doing what it says it's doing. A lot of the text models, they're able to say things very confidently, like some politicians that may not be right, but they're going to say it anyways. And so doing your due diligence to make sure that what's coming out of it is accurate is really, really great. It's great for puff pieces and things like that. I think it's going to have a big impact on marketing, for sure, and I think it already has even prior to ChatGPT, because things like Jasper and other tools like that, that will assist marketers to generate content very, very fast.

But for the world that we're in... I guess there is one other thing that comes to mind where I have seen some interesting practical examples of, let's say, AI and ML, although not ChatGPT, is things like drug discovery in the life sciences industry. And I know for a couple years now there's been work from some of the governing bodies like the FDA to figure out how do I regulate a model that I only have limited insight into because it operates in a black box. But drug discovery with AI I think is a fascinating topic as well.

James Burnand:

I was at the ISP conference two years ago, and that was a topic, and I certainly get the pharmaceutical engineering magazines, and it's definitely been something that has been a hot topic because it does create so much value for these organizations to be able to go through that process and leverage the power of AI to do that. So I'm not super involved in and aware exactly where the lines have been drawn, but being able to create boundaries for those models and methods for going through qualification of those models has been something that is a hot topic because there is a ton of value behind those outcomes.

Joseph Dolivo:

Totally. I think in the meantime, I'm going to keep generating silly images like a moose riding on a surfboard in the ocean just because I can, until something practical comes out of that. But it's a cool technology to play around with, for sure.

James Burnand:

Again, and as we look at how things in the industrial space have progressed and will progress, I think to your point earlier it makes a ton of sense is that none of this can happen without organized contextualized data. And really having a strong data strategy and connectivity to all of your sources in a consistent way, it's step one before you even think about, hey, I want to have predictive maintenance, or I'd love to use anomaly detection. You really need to have a cohesive data strategy before you even walk down that road, and maybe that's the reason why you build your cohesive data strategy, but without it, it's just a waste of time.

Joseph Dolivo:

Totally agreed. Well said.

James Burnand:

Cool. Well, I think we're probably at about time. We like to keep these things relatively short. Should anyone want to follow up with us and maybe send us some of their cool generated images, we'd love to see them. We'll definitely post some of ours with the link to this podcast.

Joseph Dolivo:

Awesome. Thanks everybody.

James Burnand:

All right, cheers.

Subscribe

Video: https://www.youtube.com/watch?v=-gyENbvvFfc&t=12s

Audio Podcast: https://spotifyanchor-web.app.link/e/gzLK3xcPMsb

Blog: https://www.4ir.cloud/blog

‍