 So suppose a delegate from Mars comes to the boss and he's hungry and He would like to eat food, but he doesn't know what food is good for him So we're going to actually tell him what is good and what is bad what is healthy food and what is what kind of food? Well, this is good Actually will kill you right so in machine learning means learning from examples positives and negative examples and in order to do so what we would like Actually to build a neuro camera the neuro camera will face objects and will decide whether it's positive or negative Edible or poisonous in order to do so we're going to borrow actually from our own neural system So here's a schematic picture of a neuron it has dendrites that get in forget information and agglomerate it toward the cell body the cell body some the information and Form some sort of non-linear operation and outputs it through the axon and the information Actually flows bottom up. So this is actually the direction of the information. Here's an artificial neuron It looks almost like a real neuron in this case. There are three dendrites You get an input these inputs are multiplied by three weights Then they are summed and go through the cell body through a non-linear function Let me do it again in pictures so that you don't have to follow the map So think of the inputs as three bars the width are the values Then the weights are three bars the heights are the values Then when we take the product we get three rectangles the area of the rectangles is the product We sum them together and we get the yellow rectangle the last operation Which is very important is the non-linear transformation Which is the red body and this is really the essence the building block of neural networks Now we're not confined to a single neuron in this case We have five inputs and then these inputs go to three different neurons that constitute what we call a hidden layer Okay, each of these neuron perform, you know its operation and outputs is transferred to the next neuron Which is at the top now the question is are we confined to one hidden layer or to a single neuron and the answer to this Question is of course. No in the following example I'm going to show a for a depth for neural network with five inputs and four outputs And again the information goes from bottom to top. Okay, and this is called the inference procedure Now once we made a cap prediction, right? We need to learn we need to contrast it with some Information so we get a critique. These are the positive and negative examples and there is naturally divergence between what the network predicted and What is actually the truth and the result is what we call an error vector And the error vector is what we're going to use in order to learn now this slide is complicated But just think about that the information flows from below the error comes from above and then based on a Hebbian rule After Donald have we're going to modify the weights Okay, and this really conjure the process of deep learning. We first go up We get a critic and then we propagate they are down update the weights and this is the end of and by now You actually know how to do deep learning and you can go and build self-driving cars So going back to our picture Suppose you actually are faced with blueberries, right? And you'd like to know what to do for images We build special networks called con of nets which are based on locally tuned neurons that further passed to Regular network and then the output is this case, you know, it's edible. Bloopers are good for you kid. Here's one Example that actually is a for it to professor soon stock. This is using artificial neural networks to detect the structure of neurons of biological neurons Let me conclude with an example of Toy example of a self-driving car on the bottom left. You see a picture. This is the camera that the neural network sees okay, and Just follow the green line Based on the image the neural network decides whether to stay straight or veer to the left and in this example We actually see that it identifies the obstacle verse to the left and then continue to the right now It all looks very simple, but there's plenty of stuff underneath. There is a basic research and massive data sets Computation storage and networking. These are the algorithms. I'm the concrete man. I'm actually building trying to build the basic algorithm Thank you very much