Introduction to Machine Learning and Artificial Intelligence

The following blog is a transcription of the presentation done by Prof. Chris Tucci during the Innovation Leaders 2018 evening. You can also see the full video below.

Prof Chris Tucci introducing the machine learning and artificial intelligence.

Thank you very much Tilo and welcome to all of you. I’m pleased that you are here tonight. We are going to talk a little bit about machine learning. Let me say a few introductory remarks and definitional type things so that way the speakers that follow don’t have to go over and over again the definitions of everything. And we can start building and finding out how people are using machine learning in business applications.

Let’s start with artificial intelligence. If you do a google search on google image search for artificial intelligence, you find an incredibly terrifying image, and I think that after tonight you’ll probably be less terrified. The general public appears to be terrified about this, and I think that will demystify it a little bit tonight.

So the goal of artificial intelligence is to develop machines that behave as if they were intelligent. That is the definition that John McCarthy developed in the 1950s and a lot of times we still think of it as a very visionary statement about AI. However what most people and most companies that are working on AI are doing is something much more narrow. That is what we call machine learning, and in a subset of machine learning, it’s a supervised machine learning by training past data
on inputs and outputs. I’m going to explain that a little bit more detail in a second.
So machine learning here is concerned with methods of learning from data and making predictions on data. Machine learning is about category labelling.
The crazy idea is that AI will wipe out everybody’s job. I went to a great talk the other day, done by Cassy Costigan from Google and she said: substitute category labelling whenever you hear the word machine learning or AI and test yourself on how that sounds: “So category labelling is gonna wipe out everybody’s job”. An interesting thought process, right?

Now imagine you have a stream of data, it doesn’t have to be a firehose of data, it could be occasional data that you have based whatever it is that comes in. All of this machine learning is based on data that comes in every once in a while that you need to analyse and you need to make some kind of decision. So imagine that you have photos your customers have taken or you have credit card transactions for your company, or you have judicial cases for crimes that need to have sentencing. Next, we analyse the data to classify these various inputs. For example the fraudulent credit card transactions versus legitimate ones. So you have a transaction that comes in as someone just made, and then you have to decide relatively quickly: it’s okay or not.

So there are certain things that you might imagine how you might make decisions on this. But there’s a bunch of things that lead up to that decision and certain
characteristics of those data set that might lead you to go ahead and say that’s a good one or no that’s a bad one. But it could also be about more complex decisions like, for example, the length of the sentence that fits the crime, tagging friends and photos. So what you want to do is go through some past data that you have, and you want to manually tag them for the decision that you’re trying to make or for the output that you want. So you could go through a bunch of card transaction say yes that one was fraudulent, that one no, that was good and so on. It is what we call making your training dataset.
In some sense tagging photos is a great example. In the photo app on your laptop, you can see how the app is trying to make suggestions about people whom you know who are in the photos. For example, it doesn’t know who this is and it’s waiting for me to suggest as to who that is. Then based on the analysis of the characteristics of the person’s facial features next time I’m taking a photo of the same person, it will tag it by default.

There are lots of different methods of categorising these kinds of decisions that you’re making so on this little graphic here you’re going to see some red X’s and green o’s and what you’re trying to do is distinguish between the red X’s and the green o’s. All the buzzwords that you hear are typically just different ways of trying to draw the line or draw the lines for the classification. So you can do it with a straight line or do you need to maybe partition it into several different areas.
By partition it in different ways, you narrow down the set, and of course, you’re never going to make it perfect, but you try and make as many good ones as you possibly can. The idea here is that the different methods that you hear about from supervised machine learning are all simply different ways of trying to draw the lines between these various sets including future data that you don’t know about yet. So usually what you do is splitting your initial data in two: Let’s say 30 percent of your data you used to figure out how you want to draw your lines. Then you bring back the remaining data and see how well your system does in predicting whatever it is that you tried to predict. At this point, if it looks good, you continue. But you need to assess every once in a while that the system is still accurate.
So once you have your categorizations and you think your predictions are working then you can start with the scale part of it. This means start handling larger volumes of predictions. You probably still need human intervention because you have to check it out once in a while to see how well you’re doing.
A great example for this is the spam detection. It’s not like you make your decisions and you figure out what’s spam versus not spam, and it works forever right. Bear in mind that the people who make spam are trying to subvert the system as we go so you’re always going to have to keep checking on your system and recalibrating every once in a while.

Before seeing some exciting business applications of learning machine, I will give you some ideas of how this is being used in business from a bigger level. Currently, the global market is estimated at around 650 million Swiss francs. The predictions about how big will be it goes from the most conservative estimate saying that within the next five to eight years it’s going to be thirty-nine billion and the most outrageous claims that go up to trillion dollars or francs. The common point here is that the market will grow rapidly even at the conservative end. If we were to look at AI in business what I would say is that the biggest gain so far has been the pattern recognition: for example voice and speech recognition like Siri or Alexa. We say when you go into someone’s house you don’t know very well you should always say: “Alexa order me some pizza” just to see you if they have something listening to you all the time. Another example is the face recognition for tagging your friends or computer vision for factories, for parts that are oriented one way or another. Good progress is done right now also in diagnostics and problem solving, so there’s inventory control retailers, financial trading bots, things like fraud detection for credit cards that I’ve mentioned already, money laundering detection, medical detection and reviewing contracts like loans and loan documents. Shortly I would say you’ll see also being used in pharma drug development, augmented reality and virtual reality.

Most people are saying “Oh! It’s going to wipe out all the businesses”. I don’t agree with this 100% because if you think about all of these applications, most of them are process innovations. In other words, these are innovations that are going to make current products and services more profitable by being more efficient through cost-cutting and re-engineering business processes. The credit card fraud detection is a great example. It’s not going to wipe out the credit card. In fact, it’s going to make credit card companies more profitable because they don’t waste so much money paying on merchants for fraudulent transactions. The same phenomenon happens with the computer vision, same thing with manufacturing diagnostics data.

Just a couple points here about AI business models: In other words, what do you do with these data? How does that affect your product service offering? Companies are coming up with a little bit more servitization bundling services, paying for uptime IP servitization and monetizing data.
Let me close with this one quick example which is a very interesting product that I’ve heard of recently. It is a pill that you take and after you swallow it, it gives off data about the state of your inside. So then they analyze data they can compare with healthy people, and not healthy people. They can do lots of things with this, but the question is how does that change their business model? Should they be selling the pill to insurance companies? Should they be selling up time for patients?

Let me thank you and welcome you once again I hope you have a great night.