Multi-Layer Perceptrons

Multi-Layer Perceptrons (MLPs) were a big innovation in the 80s.
A perceptron just takes the inputs, multiplies them by weights, passes them through a function, and sends out the output.
So, they're computationally really simple, and they're closely related to McCullouch Pitts neurons.
You can use any function, though a sigmoid is common. It's a lot closer to a discrete McCullouch Pitts neurons if you use a step function.
You can have a layer of these, where they all have the same inputs, but have different weights, but don't get much more power.
If you have multiple layers (typically two), you get a multi-layer perceptron. This can approximate to an arbitrary degree of precision a huge range of functions.
The cool thing is that using a back propagation of error function, you can learn to set the weights for training data.
If you do it right, it generalises to testing data.