 Okay, so, yes, thanks a million to Edward and Ricardo for what is a very timely and appropriate talk to go before me, because what I want to talk to you a little bit about today has many themes that are similar to what Ricardo was mentioning there. On one hand, we have projects, interesting projects that we do within the CDAR Center that build on the state of the art in machine learning, data analytics, etc. But importantly, not all of these projects have to be about banking, have to be about finance, have to be about many of the bread and butter issues we've been doing very successfully for many years. They can be in different areas. And when we were talking to the organizers of the Predict Conference, we thought one thing that might be interesting to really focus in on this year was the issue of sustainability, the environment, and how we as data scientists, all of us here as practitioners, can try to aid and help to move the world we live in in a good direction where AI is being used for good, essentially. So that's a large aspect of what we do, particularly in terms of the individual projects that we execute in CDAR. So I'll say a little bit about some of those projects and some of the techniques we make use of at the moment. So I am a researcher in CDAR and also based in TU Dublin, which is formally known as DIT for those who didn't notice the name change earlier in the year. And as part of CDAR, what we do are these very directed industry projects. And a lot of these that I'm involved in in particular are in the agri-food space. So you're probably sitting there saying to yourself, why agri-food? Why the hell would anybody care about agri-food one way or the other in an AI venue? Well, the answer to that is quite simple in a way. It's important. Let's consider the statistics on where global population and food consumption is going in the next 40, 50 years. If the problems haven't gone away, even though we hear about Europe and Japan having very much an aging population, the global population as a whole is continuing to boom with estimates of us hitting the whole 10 billion mark by the end of the century. And that's even based on the conservative estimates. So that's definitely where we're going in terms of a jump in population. But importantly, it's not just that we have a jump in population. The problem is that the population as it expands, it's eating different types of food. So we have the consumption of meat and of dairy and all of these very protein intensive foods really taking a jump as well in terms of our predictions. And we need to be able to cope with that increasing demand. And that's what has given rise to this area of precision agriculture, but also precision food, precision food processing, which is just as important as the work that's actually happening on the farm itself. So what sort of challenges do we face in this area? Well, for a start, even though the population is increasing and the actual consumption of individual food types is increasing, we have to balance the work we want to do in feeding the population with the very fact that disease and over farming are very much challenges we need to put up with. So the question starts to come. How can we try to help farmers? How can we try to help food producers to optimize what they produce without having long term damaging effects on the environment itself? So ultimately we want to help to keep sustainability there. So that's one of the big challenges we face. But the challenges we face aren't simply out in the field itself. If we look at the food production side of things, we can try to look at what is essentially industry, what is essentially a mass factory environment, and try to aid production within those environments as well, either by looking for problems in a process or maybe assisting human workers to do work faster. Again, we need to be able to help with these things. So what can technology, what can AI start to do for us? Well, fortunately, there's many different techniques from a computer vision, from a machine learning type process that we can leverage. And here I've highlighted some of the more important ideas. Up at the top you have that idea that the world isn't just what we perceive with our own eyes. We look at the world in a variant of red, green, and blue, but that's only part of the complete electromagnetic spectrum. So a lot of practical work we do in CEDAR projects, in projects with Irish SMEs looking at food and agriculture, is trying to make use of the full electromagnetic spectrum and applying ideas from computer vision to that full electromagnetic spectrum. Because as you can see from just the image I use here for illustration, that full EM spectrum can give us an awful lot more information on what's actually going on. So we want to try to leverage it. And fortunately, we have the power of deep learning, computer vision, instant segmentation, very deep, pre-trained networks that we can leverage for all of that work. And I know people in the audience know the bread and butter of these techniques very well already, but it's nice to see them being used in novel and interesting ways. So overall, we can use these technologies to try to do very practical tasks like monitor animal health, to try to limit the amount of fertilizers that are being deployed in the fields, limit the amount of weed killer being deployed, etc. All of these are very important goals. But there are big challenges. And what I think is interesting is that many of the challenges we face in these agri-food type projects are actually challenges faced by many different, not just computer vision, but deep learning AI type problems in general, including, for example, the topics Ricardo was talking about. And the challenges in particular that I see are just pure collection of data. If we're all working with tweets and we've got massive tweets, data sets, it's pretty easy to do some interesting work. But when we're working in new and novel domains, one of the biggest challenges we have is just collecting relevant data. We have the challenge that even if we are able to collect some data, it's often not nearly enough to truly take advantage of deep learning models in the way that deep learning was intended to be used. So that's also a very significant challenge. And we have, particularly for my domain, but it's also true in other very practical real world domains, that we don't have necessarily the access to the cluster, to the back end data center to do all of the number crunching. So how do we cope with that sort of limitation? Can we develop models that are still interesting and powerful and deploy them to more limited domains? So just in the time I have left, I want to talk through a little bit of a case study on a project that we are doing at the moment and talk about maybe some of the ways in which we've tried to tackle those problems. So this is a collaboration we are doing with an Irish SME called Tanko, who are based down in Carlow. And this is, as the guys were talking about earlier, one of these projects that's semi-funded by Enterprise Ireland, also funded by the partner company as well. And I have to thank my partner on that project, the CEO of Tanko, for some input on the slides that we're using today, because it is a case study on there, our joint work with them. In a nutshell, what we're trying to do with this project is to take advantage of the state of the art in machine learning, in computer vision, and apply it through these multi-spectral imaging methods to try to do interesting things around improving the quality of grass, essentially, and how grass is used to produce silage. I know that doesn't sound particularly exciting for all of us cement dwellers, but it is pretty important in terms of making sure that farmers are getting value for money in terms of the work they're doing, putting into their yields, et cetera. So what we've tried to do with that project, still ongoing, we're still working away, is a lot about data collection. How do we gather and build data sets that are relevant? And for us, this has meant very complicated engineering tasks in terms of putting sensors together, lots of different wave bands, other sensors I don't want to go into for IP and protection issues, but some really interesting engineering, but that's ultimately coming back in a data set that we can apply with these state-of-the-art methods. But in terms of the bigger challenges, I just want to say one or two or three things about where I think people should go and have a look at interesting ideas if you want to go out and be inspired to learn something by the end of this talk. One thing to consider is while we can do a sort of a roll-your-own approach to the network design, we are of course taking advantage of very big, deep networks. And what we found that's really interesting is that even for really complex networks like Resnet and like the other pre-trained networks, even though they're not completely domain-relevant for us, we are seeing fantastic results for them. So I guess my learning there for you is be adventurous in the number of ways you try to take these pre-trained models and put them to work. You can have surprisingly good results even if in theory you think it shouldn't work. And I always think for those who have more of an older AI background like myself as well, it's like that whole thing about the naive Bayes classifier. Even though certain rules have to be obeyed in theory, often you can try naive Bayes at a problem and it gives you pretty good results. Another thing we looked at is applying transfer learning because again, we have that challenge. You've got a small data set. How are you going to try to leverage that data set? Well, make use of ideas like transfer learning. Don't build an independent model for your problem. Build a model that's based on some big pre-trained data set that's much broader and then try to specialize it into your limited domain. That's a very important learning that we should all be making, taking advantage of. And the final one that I want people to be very aware of and to actually investigate and have a look at if you're working in computer vision is we have that challenge. How do we take really big deep networks and put them to work in limited domains? And what we're doing is taking advantage of these technologies like the Movidius' Myriad chip, which is really trying to focus down and concentrate deep learning into a really small footprint. What the guys in Movidius have done is really, really fascinating, interesting stuff. They were bought out by Intel last year or so because the work was just so interesting and relevant. So, if that whole area of deep learning is something that excites you and you want to maybe see it being used in more novel or interesting domains, have a look and investigate Movidius and the Myriad chips in particular they've put together. So, that's it. I get off the stage now. All I wanted to do is really just try to give you a flavor of some of the projects we do in the agri-food space and maybe hopefully give you some thoughts on what might be some interesting technologies such as the transfer networks, deep learning and the Movidius' chips that are really exciting for a wide range of projects. Thanks again.