 Let's dig into the artificial, I would say, intelligency world. So what we are seeing is artificial intelligency is a super set of everything concerning, I would say, related to machine, mimic, cognitive capability like humans. So each time we are trying to copy or to have a mimic of human staff, so it's in a big category of artificial intelligency. So okay, it's covering everything. Everybody is talking about artificial intelligency. It gives you possibility to have some interaction with the environment, knowledge, perception, learning and so on. These are the keywords we are learning. And this concept are coming from computer science. So you need science, mathematics, statistics. These are the three. But this is a super set. And inside the artificial intelligency, you have one of the subsets. It's called machine learning. Machine learning then means that, okay, you will collect a lot of data set in this case. And then you will start to train based on this data set to train, to extract some information in order to learn and to have some prediction in this case. So this is machine learning. So the machine will learn, I mean in this case, will use the machine to learn and to train your staff. So it could be mathematics. It could be statistics. It could be, and I will show you some examples. And inside the machine learning, what can we do? We have some pattern recognition. This is a keyword. We have some anomaly recognition, some prediction in the future. So it could be stock price in the future, advertisement and so on. Anomaly recognition. You have some machine and then someone is trying to change the neural of the machine. You can recognize it. So this was, okay, bank are doing and other stuff. And also pattern recognition trying to spark more facial identity and so on. So that was the key. You have three different kind of learning stuff in the machine learning. So you have the supervised, the unsupervised and reinforcement. It's a three technique they have. One of the technique, and this is a one, is a supervised. So that means you will have data which are labeled. What does it mean? So if you want to recognize, I don't know difference between the different animals. So we'll put picture with cat and you will label it as cat. You will put picture with dog or with other animal. And then you will label all this data. So big, big job to be done. So that means you have to pre-prepare this data. And then you will start to have a learning process. So you will learn from it, from this data. And then you will try to base on different algorithm, base on different technique. I will explain this different technique after all. And the goal of this after this learning is to recognize. So this is one of the method. Second method is we say reinforcement. So you are showing stuff and like I would say dressing a dog. So if it's good, I mean if the result is good then you are giving in a sugar or something which is good. And otherwise it's a kind of rewarding system. So we are rewarding and you are learning. So your system is learning based on that. And the third one is unsupervised. So there is no control. There is no label. So we are trying to on the fly to have some category and then after all to decide what to do. The good thing is you can mix of this. So these are the big models. So this is triangle but the learning is something in between. So you can use part of it and then mix. Okay. But this is a general stuff for machine learning. And within this machine learning what we have is some of the approach. You have much more approach but okay these are the main approach. I mean we found. So the first one is a decision tree learning. So decision tree. Okay we have the picture here so it's quite clear. Based on the data set you got. Then you are trying to extract some you know feature and then try to I will say the machine is trying to transfer it inside the decision tree if then else and so on. Okay. This is one of the technique. Second technique is clustering. So trying to put you know all the data inside clustering and then trying to find the area where this cluster are to predict the next stuff. So this is a method based on statistics also on mathematical stuff. We have rules based learning. So this is something different. We have inductive logic programming prologue. You can have a look on the web. So it's different kind of things mixing statistic and mathematical phenomena. And the last one and this is the one we have interest today is called deep learning. And using neural network in this case. Deep learning neural network. So this kind of stuff. And this is where we are focused. But deep learning is again a sub system from machine learning. Okay. At ST we are also using decision tree learning and clustering. So we also have algorithm on our sensor. So we have sensors for instance which are embedding some decisions we learning. So inside. So this is AI. Of course we are calling AI. This is a machine learning. But in this case it's not neural network. No we will dig it a little bit more inside this part. Deep learning neural network. Why? And this is why we are explaining this. Because what we have seen in neural network deep learning is existing things long time. It was going up and down because of you know resources we had because of the performances. The data set we are able to collect. Why? Because we have some advantages of the learning and disadvantages. What you are saying is not it's ideal for everything. So at the end what you will realize during your if you have no experience on the AI from now. So you will have to try maybe all of this. And then select you know the best for you at the tea time. And then maybe we'll change based on the amount of data you will have based on the accuracy you need and so on. So what we are not saying is okay this is the ultimate if no no this is one of them. And this is one which is quite complicated normally to port on the microcontroller for this one. You already have some we say software which are existing on the web which are converting you know this in C code. So this is existing but for this one there was from now nothing or not so much. It's the reason why we are concentrating on this one. What are the advantages of this one of the deep learning is there. Advantages is okay high accuracy and what we are saying is okay you have a performances and you have the amount of data. And at the beginning I will say let's say that the normal algorithm decisions reclustering rules based learning and other stuff are more efficient than the deep learning because deep learning is quite complicated to put in place. It's requesting high computation and so on. So these are the disadvantage but at the beginning okay you will see advantages of all the algorithm and then there is a breaking event. So it depends on the application question is when is happening okay it depends on the amount of applications depend on the complexity of your application. But then you have a break even and then the deep learning is continuing you know to bring you some performances where the standard learning algorithm has stopping somehow. Why stopping because with the decision 3 it's a amount of you know memory space you have in this case or somehow each time for instance you're getting new data you will change completely the structure of the 3. So it could be difficult to qualify in some case and depend on what you have and so on. So these advantages is you need a large data set to start. So when I say large it's really large I mean it's a big amount of data to record. These advantages is a weak theoretical explanation. So it still looks like a black box in most of the case. It's running of a customer we were walking five years long on the algorithm trying to detect some some some characteristic doing some kind of predictive maintenance based on the audio spend five years you develop algorithm based on mathematical stuff. And at the end they get something which were almost walking let's say so it was a walking was recognizing what they want and then some condition change in the room. And then everything was gone so all the you know the five years was not working anymore so they had to read change and all the stuff and OK it was a lot of job. And then they engage to AI guys and to try to use a neural network in this case and this guy had a success after two months really first success. And it was so successful that successfully that OK within the two months achieved to have at least at the same level at the you know the five years development before. And more because then there was not in big impact about the change in the environment so they are really. But the big issue they had was to explain it because it was safety based in this case. So they had to explain it to the fine customer and it was not able. So they had a good algorithm which we are working well but there were you know problem to explain why it was working. So good in this case. Yeah it could be the problem. So this is a disadvantage is we have to put it also in the balance. So it's what we are saying is not of use that OK this will be you know the right algorithm for you in each time and each. It's a reason why most of the people in the partner are talking about machine learning in this case because they are proposing you know different algorithm which is OK in this case. Also comparison OK to be honest these are the problems I mean in the AI since the beginning. So problem of object classification. So which are object we are showing some object classification in this case send classification audio send or other stuff object detection. You are putting some 1000 objects in a different object or more and then you have to detect it and then you have some data set which are on the web. So you can download all these big data set and then you can leave your algorithm running and these are the best accuracy you got without using the conversional. I mean we say neural network in this case so and OK for some of them it was quite good. I mean in this case face recognition 96.3 so quite quite good with conventional stuff for some other it was OK. And on the rest on the right hand side you have the best accuracy using the neural network and then you see some plus in some case you are some plus big plus. And some case not big plus so it's not obvious again. So it's really depending on what you are doing what you plan to do. This is a message but it's some cases broke so much that some of the people said OK we'll not use any more conversional stuff depending on the you know the problem you had. We'll use no CNN and they are focusing on the neural network stuff.