Скачать или смотреть Multilayer perceptron

Multilayer perceptron is a classic feed-forward artificial neural network. It is one of the core component of some deep learning algorithms. Not all algorithms in deep learning use a feed-forward artificial neural network, but many do. Another way to look at this, is that a multilayer perceptron is a connected series of nodes, where each node represents a function. This connected series of node, create what is called a directed acyclic graph. Meaning that there is directionality between the nodes and no node will ever be revisited.
 
This all starts with an input layer. Let's use the Titanic data set example. There are four nodes here and each node would represent a feature that we want to use to predict whether somebody survives or not. These could be age, ticket class, cabin and sex. Next there's a hidden layer that we're representing with five nodes here. But this hidden layer could contain 10 nodes, a hundred nodes, or even a thousand nodes. Now each of these nodes represents some function. What happens is, we take each of our input nodes, representing features and we pipe those into each node in the hidden layer. What's happening here is each node, or function, in the hidden layer is getting each input feature. So age, ticket class, cabin and sex are passed into each node in the hidden layer. You could view each of these nodes as something like logistic regression. So, more or less a logistic regression model would be fit for each node made with slightly different weights.
 
Then we have an output layer. And here we're going to have two nodes. Either a person survived, or they did not survive, using the example of our Titanic data set. Now we have each node in the hidden layer, outputting to each node in the output layer.
 
Let's take a step back and just isolate one node in the hidden layer to understand what's happening here. You can see it's receiving the four inputs. Then this one node in the hidden layer will basically fit some model, or learn some aspect of the data. Then based on that model that it learns, it'll make some prediction about whether somebody survived or not. Maybe it's 70% likelihood that they survived, 30% likelihood that they did not. That's just one individual node in the hidden layer. The same process can repeat for each of the five nodes in the hidden layer, and you get a final prediction of the likelihood of a given person surviving.
 
As you can imagine, these can be incredibly powerful. As you see in this example, you're essentially aggregating the predictions of five different models all together. This allows the overall model, to learn some really powerful relationships in the data.