Skip to content

Page394

Real Neural Networks

Let us discuss how a real neural network operates: Imagine you are walking down the street in a city at night, and someone is walking behind you closely. You begin to become nervous: it is late; it is dark; and the person behind you is too close. You must make a decision: fight or flight. You must decide to turn around to face your pursuer, or to get away from them.

As you are making your decision, you weigh thousands upon thousands of inputs. You remember past experience; your instincts guide you, and you perceive the world with your five senses. These senses are sending new input to your brain, millisecond by millisecond. Your memory, instincts, sight, smell, hearing, etc., all continually send synaptic input to neurons. Less important input (such as taste in this case) has a lower synaptic weight. More important input (such as sound) has a higher synaptic weight. Neurons that receive higher input are more likely to fire, and the output neuron eventually fires (makes a decision).

Finally, you decide to turn and face your pursuer, and you are relieved to see it was a person listening to music on headphones, not paying attention to their surroundings. Thousands of inputs resulted in a binary decision: fight or flight. ANNs seek to replicate this complex decision-making process.

How Artificial Neural Networks Operate

ANNs seek to replicate the capabilities of biological neural networks. A node is used to describe an artificial neuron. Like its biologic counterpart, these nodes receive input from synapses and send output when a weight is exceeded. Single-layer ANNs have one layer of input nodes; multilayer ANNs have multiple layers of nodes, including hidden nodes, as shown in Fig. 9.13. The arrows in Fig. 9.13 represent the synaptic weights. Both single and multilayer artificial neural networks eventually trigger an output node to fire: this output node makes the decision.

Fig. 9.13 Multi-layer artificial neural network.

An Artificial Neural Network learns by example via a training function: synaptic weights are changed via an iterative process, until the output node fires correctly for a given set of inputs. Artificial Neural Networks are used for “fuzzy” solutions, where exactness is not always required (or possible), such as predicting the weather.