2 Neural Networks

Now that we have defined a single artificial neuron, we are ready to place these neurons into a network.

2.1 Networks of neurons

A single neuron is rather useless unless it is connected to many others. This is true of both biological and artificial neurons. A Neural Network (NN) is created when the outputs of neurons are used as inputs to other neurons. Clearly the whole network also needs some input/output interaction with its environment. This is achieved by having some neuron inputs come from the environment, and some neuron outputs being used by the environment. In this way the NN performs some processing of inputs (or "measurements") from its environment, and delivers a result (or "commands") back to its environment.

2.2 Feed-forward networks

Although it is possible to connect the neurons in many different ways, ANNs are often arranged in layers of neurons. The outputs from each layer are connected to the inputs of the next layer, and there is no interconnection between neurons within a layer. This arrangement is often called a feed-forward network because signals always travel forward from the inputs to the outputs; there is no feedback.

A typical feed-foward ANN.

One neuron illustrated.

Each neuron is shown as a trapezoid with many inputs and a single output. The large circles on the left represent distribution points from which the external inputs are distributed to the first layer of neurons. Every neuron in the first layer receives a signal from every input. The input is multiplied by a weight as described earlier. The bias terms, which were not shown in above, are shown here as if they were weights for an additional input signal with the constant value of 1.0. This allows each bias to be treated as just one additional weight.

Just as each external input goes to every 1st layer neuron, so each output from these neurons goes to all of the 2nd layer neurons, where it is multiplied by a weight. And similarly the 2nd layer outputs go to the 3rd layer inputs. The 3rd layer neurons produce the output of the ANN. There is exactly one 3rd layer neuron for each required output.

Note that it is often ambiguous as to whether a layer is a layer of nodes, or a layer of the weights between the nodes. As both uses are common, beware!

Back to top. >>>

Continue with the tutorial. >>>