AI applications, machine learning, human-machine interaction

Thank you very much. So what are the main AI technologies and the main AI applications up to now? So, AI is already used in our daily lives. If you think about search engines, Google or computer vision used to recognize objects, or code bars on your cell phone, or speech recognition, anything that has to do with complex problems that require some level of intelligence to be solved. And how about object recognition and voice recognition? Can relate face and speech recognition? Yes. So anything that has to do with recognition of objects on pictures or on video streams is really part of this large field called computer vision. And you can recognize objects, in which case you sort of define the objects that you want to recognize. Like you want to recognize couches or different kind of furniture or Legos or whatever to define the categories that you want to recognize. Typically, you have a system that trains a model to recognize these kinds of objects. And for face recognition, it’s the same, except that instead of physical objects, you work as human faces. And then again, you need a system that you’re training a model that recognizes different kind of persons according to their face. And for speech recognition, it’s essentially the same, except of course, it’s not a visual input, it’s an audio input. But the idea is the same, that you recognize some patterns in the audio, in the audio data that defines different kind of sounds and putting together the sounds, they will make words, they will make sentences, and then you can recognize what is being said and transcribe it. Could you tell us what are the emerging trends of today? AI is a large field already, and it’s growing at a quite fast pace. But some of the biggest trends, of course, is given the widespread availability of large amounts of training data. One of the important trend is really to scale up the technology so that it can process really large amounts of data, everything that you can find on the Web, basically. And that really is both for language technology and so text and computer vision, because you have billions of images on the web. So this scaling up technology is one important trend. On the other hand, an important trend which I find very interesting is really the fact that more and more researchers are looking at the ethical issues behind the technology. Partly because, well, for instance, we’re using basically everything that is found on the Web without tasking human persons that are shown pictures for their consent about their usage for these large models, but also because we realize that AI has really an influence on our daily lives. And then we should think about whether, for instance, AI can become biased or express stereotypes or become hateful, sometimes even because it’s trained on data that is found everywhere on the Web. And we know that the Web is a strange place sometimes. So thinking about how AI what kind of what does the model learn when they scoring the web for data and how we can make sure that they meet our expectations? And, of course, regulations when it comes to non-discrimination. For instance.

And what is the difference between machine learning and deep learning? So deep learning is a particular kind of technology for machine learning. It’s not the only one, but it’s become a very popular one for some types of problems. The point there in deep learning is to learn so called neural network networks with many processing layers. So you have a quite complex mathematical model which is comprised of small computing nodes that receive information and then forwarded on to other neurons or other computing nodes that have specific mathematical properties. And even though these small nodes are quite simple in isolation, the fact that you combine them together and connect them in large networks with millions of such notes, a bit like in your brain where neurons are connected to one another, can make systems learn how to perform complicated tasks. For instance, deep neural networks. And deep learning has been used for machine translation. So the automated translation like in Google Translate, we know that it’s a complicated task, machine translation, you have to understand the context of the sentence. You have to understand the linguistic properties of both the input language in the target language. You have to understand how words combine with each other so that if two words are put together, it makes construction that is translated differently. So these are really complex tasks and we know that using traditional machine learning Latif was stored until 2010 that give results that were okay most cases, but with serious mistakes and problems understanding contextual factors, for instance, whereas these deep learning networks were better shown that they could find translation of high quality. This being said, it’s far from an open problem and there are still many problems with machine translation but it has been shown that they were better at understanding how to translate. And what is the difference between machine learning, deep learning and traditional programming? So machine learning is really an umbrella term for everything in AI that is trained from data. So it’s really quite large. It includes deep learning but also with other techniques, some of them dating back to the seventies. While deep learning is really a specific kind of technology based on neural networks which works in many cases, but there are many other techniques that are working well and have other properties that are interesting. Maybe I can say that one of the main shortcomings of neural networks. So first of all, it requires large amounts of data and for some problems you have large volumes of data base far from the case for everything if you translate it. For instance, machine translation for a language for which you don’t have a lot of resource, it’s going to be difficult to use deep learning. Another problem is that these are complete black boxes. That means you don’t understand what the system has learned and sometimes it’s okay. You don’t always need to understand everything. But if, let’s say you’re building a system for deciding whether you should give a loan to a person. And you have a system that may based on the salary of the person and where they live and whether they married or not, et cetera, et cetera. Different kind of demographic factors. And you have a system that may give a good prediction about whether a person is entitled to no, no. But it’s not able to explain the rationale behind their decision. It’s going to be heavily problematic because you need a system that not only gives an answer but explains why they came to that particular answer. And so that’s an important shortcoming of neural networks. And that’s why sometimes machine learning systems may have a slightly lower performance on some respect but are able to explain the steps, the processing steps behind their decision sometimes are much better. So there’s no one single way to evaluate those models. There are different kinds of pros and cons behind different kinds of technologies. Yes, And traditional programming is simply what you do when you don’t learn the model, where you simply pre-program all the rules in advance.

Now we will move to human mission interaction. What are the goals of human machine interactions, in your opinion? So the goal of human machine interaction is simply to create good interfaces between a human individual that wishes to perform a given task and a machine that supports that task. And just a graphical user interface isa very simple example of this. But of course, anything that has to do with interaction, with technology is part of this. So I’ve been working on talking robots, for instance, and that’s another kind of interaction that is quite different from a window on your screen, but it’s also an interface in the sense that you in that case using spoken language to provide instructions to receive feedback on what’s going on in the execution of the task. And the question is always how we can have a communication that is fluid, where the person understand what’s going on and can easily convey what ought to be done. And how can devices be controlled? It really depends on the technology. I mean, for a computer window it’s rather straight forward as long as you understand the interface that a small cross means that you closing the window. And for more sophisticated tools, it’s of course a bit harder because you need to find something that is intuitive enough, yet also powerful enough to enable the range of possibilities for the particular device that you have. For a robot, for instance, it’s quite large, but the set of possible things that the robot can say and do in the world. So you need some kind of and it’s one of the core tenets of human robot interaction, you need some kind of transparency. That means the system should as much as possible provide very explicit signals as to what the system has understood, what it has not understood, where itis in the execution of the task, so that the user is as much as possible in control, because without information there is no control.