Wie funktioniert eigentlich Machine Learning?
The ads you see, which price you pay, which movies get recommended and even who your next date is. All those things today are influenced by artificial intelligence. Most people have heard of AI before. It’s a very hot topic right now. But many people might not know how it actually works. How do algorithms that can seemingly learn anything, work? The term ‘artifical intelligence’ is a bit dramatic and often used in popular science. Experts speak of Machine Learning instead. But how does it acutally work? How can a computer that works solely with 1s and 0s, learn something? When I ask you which of these two pictures shows a human, then this should be no problem for you. We can also identify blurry, covered or hidden people. And even a new concept that you’ve never seen before can be learned in seconds and recognized without me telling you the specific elements that make up this concept. But when it comes to explaining how we can recognize something, we are clueless. How do wo learn? Well, of course learning hast to happen somewhere in the brain and it therefore has to do something with the neurons and synapses In our brains. But while we understand how single a neuron works it is nearly impossible to predict how many interconnected neurons work together. Dr. Jascha Ulrich: One approach of machine learning are neural networks. The idea of neural networks a loosely inspired by how the brain might work. Moderator: This is Dr. Jascha Ulrich who works at ZEISS’s ‘active research’ department on artificial intelligence and I asked him to to explain to me what neural networks are in a nutshell. Dr. Jascha Ulrich: Ok, very simply. For a neural network you need a neuron. A neuron is simple, it get’s an input and gives you an output. The output can be different from the input. It’s very simple. Now I can connect multiple neurons together to form a very big network. And now I can accomplish much more complex tasks with a lot of little pieces that all do simple tasks. Moderator: To get a better idea on how this process works let’s build our own neural network. Let’s say we want it to recognize faces. At the beginning our network is nothing more than computer program that simulates single knots. Those knots are really called neurons. And one neuron only has one task: to receive a number, alter it in some way and to send out the answer. For example: a neuron could have the task to half every number it gets. Also, our network is organized in layers and can send values from one layer to another. At the end there are only two neurons: “face” and “no face” which are our result. If everything’s working then of those neurons should display a high number when feeding a picture into the network. Because we started with random values the result in nonsense. The network is total chaos and doesn’t make sense. And that’s ok because we don’t want to tell the network how to work. Now it get’s interesting. Now our network is going to learn. A neural network is a mathematical function. And by tweaking the values of this function out network is able to “learn”. When we feed a picture into our network and receive a random result we can push our network in the right direction. We know the picture contains a human face and so we want the neuron “face” to receive a high number and the neuron “no face” a small number. To achieve this we can influence the connections to our neurons by altering the values in the layer before that. And to alter those values, we need to change the ones in the layer before that. Every value matters and changes the result because everything in our network is connected. After we fed the first picture into the network and tweaked a tiny bit we feed a second picture and then a third. And so on. With every single picture we feed into the network we look at the result and tweak the hundreds, thousands or millions of values so that the result is a little bit more correct. Depending on the task you need millions of training pictures that all need to be labeled by humans. This is a lot of data. But it’s worth it. Because our network is now able to distinguish between pictures of faces and pictures without faces.