Neural networks are an important approach in machine learning. These networks are composed of multiple but simple processors that work in parallel to model (non-linear) systems, where there is a complex relationship between input and output. Analogous to our brains, the processors in a neural network are also called neurons.
A neural network usually consists of multiple layers: an input layer of neurons representing the input of a problem, an output layer of neurons representing the solution of the problem, and intermediate layers with artificial neurons that perform calculations. Each connection can transfer a signal to another neuron. Neurons have a weight that can increase or decrease the power of the transmitted signal. In other words, neurons are activated via weighed connections of previously active neurons. The receiving neuron processes the signal and then sends a signal to the next neurons. The threshold is important here: Only if the aggregate of the signal is lower (or higher) than that threshold level is the signal transmitted.
The interaction between the processors in a neural network is adaptive, so that connections between other processors in the neural network can be formed, and existing connections can be strengthened, weakened or broken again. This means that a neural network can learn. Here, ‘learning’ refers to the automatic adjustment of the parameters of the system, so that the system can generate the correct output for a given input.
Information that flows through the network has consequences for the structure of the neural network because it changes – or learns – on the basis of that input and output. Take the example of pattern recognition: Neural networks can identify images of cats by analyzing sample pictures that are manually tagged as ‘cat’ or ‘no cat’ and use these results to identify cats in other images. Neural networks are most useful in this type of applications, which are difficult to express in a traditional computer algorithm using rule-based programming. Applications range from optical character recognition (printed or handwritten scans to digital text) to facial recognition.
Another example: Suppose you want your algorithm to make predictions about fraudulent credit card transactions. In this example each layer of the neural network is able to recognize a specific feature. For example, the first layer looks for high amounts, the second for banks abroad, etc. If the system comes up with a wrong answer, the network adjusts the weight given to certain neurons. For example, a foreign bank may end up being less important in recognizing a fraudulent credit card transaction and thus should be given less weight. The next prediction will be more accurate in this way.
