Neural Network Simply Explained - ML for Beginners

hi everyone today we will talk about my favorite machine learning algorithm the neural network neural networks were designed to mimic the process of human thought we use them to solve problems that traditional computer programs find extremely difficult or almost impossible to solve for example face recognition object detection and image classification in all three cases humans are really good at it we don't really need to think much to know that this is maria and this is not maria we can also clearly see that there's a keyboard a mouse and a monitor on this desk and we can also say with confidence that this is a photo of a goat but if we are so good at it how come computers are not aren't they supposed to be smart or something when a computer sees an image it actually sees a collection of pixels where each pixel stores a numeric value representing the color intensity we already talked about it in previous lessons and you guys can definitely check it out if you haven't so that instead of a goat photo our computer sees a very long array of numeric values and then the tricky part is even if we take another photo of the exact same goat on the exact same background we will most likely get a brand new array of numeric values and guess what even if we use the original photo but slightly boost the brightness or maybe crop 5 pixels from the width or even just rotate it 5 degrees clockwise we still get a brand new array of numeric values that has nothing to do with the original one so if it is so difficult how can we ever classify an image and left alone recognize a face in it we simply expose our neural network to an enormous amount of images each image is an example and the more examples we provide the bigger the chance that our neural network recognizes the subject in our case that would be the class or category of goat but that's not all our examples also need to be interpreted properly that's why very often each example will have something called a label or a target that tells our neural network to which category it belongs we call this process supervised learning and if you guys want to find out more about it you can check out my previous tutorial but between the point where we feed an image to our neural network and the point where we reach the conclusion that it's a goat a lot of different processes are happening in the background these are series of statistical calculations that investigate each of our examples from many different angles these calculations happen inside something called hidden layers they are located in between our input layer and our output layer and they are responsible for evaluating different aspects of our examples for example one hidden layer is responsible for edge detection another hidden layer maps the colors another one counts the legs or maybe even detects horns now on their own these layers are quite useless just because something has horns doesn't really make it a goat but when we combine all these layers together we actually have enough information to make a prediction so an image is loaded into the input layer along with its label this is where we begin training our neural network the image will then pass through each of our hidden layers one at a time and once we reach the output layer our network will return a class name which we call a prediction this class name however will not always match the label and the main reason are those links that connect between the nodes of our layers we call these links weights and they help us determine how much impact each node has on the input for example maybe counting legs or detecting horns is much more meaningful than mapping colors in that case the weights that lead to the color layer will have lower values than the other two so just because our first prediction is incorrect it doesn't mean that our neural network is bad it only means that we need to keep adjusting the weights until most of our examples are correctly classified or predicted we call this process optimization it can take a very long time and we would usually optimize other parameters as well not just the weights but we'll talk about it in future lessons but the good news are once we are done optimizing we can save the neural network and we can load it whenever we'd like then we can expose it to goat images it has never seen before and in most cases they'll be correctly classified now because our neural network has already learned about thousands of goat images from different colors different angles and different ages when it comes to goats it's an expert but when it comes to dolphins and giraffes it is absolutely clueless that's what we call narrow ai it can only do a one particular task and it has a very narrow area of expertise it is up to us developers to keep teaching it further and making it much much smarter cool so in this video we mostly focused on images as input however we can also classify text we can classify audio video and pretty much anything else that can be represented with numbers now thank you guys so much for watching if you found this tutorial helpful please share it with the world and you can also leave me a like you can leave me a comment you can subscribe to my channel turn on the notification bell of course and i will see you guys very soon in a brand new machine learning code along

Channel: Python Simplified

Views: 6,874

Keywords: ml, ai, machine learning, artificial intenligence, artificial inteligence, artificail intelligence

Id: i1AqHG4k8mE

Length: 6min 37sec (397 seconds)

Published: Sat Oct 16 2021