Skip to content

Neural Networks

Neural networks

The so-called transformer architecture of generative AI such as GPT involves artificial neural networks (ANN), which are modelled on the functioning of natural neural networks, e.g. in a brain.

The neurons in a neural network are arranged in layers one behind the other. We speak of the input layer (red), the output layer (green) and the hidden layers (blue).

Neural network with input, hidden, and output layers

A single artificial neuron is connected to all neurons in the upstream layer. The connections should not be thought of as switches (on/off). Instead, the signals of all inputs are weighted and used as network input with a transfer function. An activation function leads to the activation of the neuron (the neuron fires), taking into account a threshold value. The weightings correspond to the parameters of the network (a Llama 2 7B model, for example, has 7 billion such parameters).

When training a neural network, you start with random parameters. In the training process, the parameters are set through a process of machine learning by calculating an error function so that the neural network provides the most correct answers possible.

This explanatory video shows how a neural network can recognise numbers (e.g. the postcode on a letter). A 28x28 pixel image is used as the input layer (784 inputs). The network used has two hidden layers. The output layer has ten outputs (indicators for the numbers 0-9).

Comments