Neural networks are groups of interconnected neurons that work together to perform a particular function. Until recently, this could only refer to biological neurons and networks, however progress in computational neuroscience has led to the widespread use of artificial neural networks, so the term can now refer to both artificial and organic circuits of neurons.

This website will focus on artificial neural networks, as this is where the computer science comes in.

Neuron diagram by Katie Dektar

Neurons, or nerve cells, are the basic units of the nervous system. They receive inputs through the dendrites, and the chemical/electrical signal is carried away onto the next cell through the axon.

Neurons communicate through synapses

Neurons form connections with each other through their synapses, across which chemicals can be released from one neuron to the next.

Artificial neural networks are constructed from artificial neurons. McCulloch and Pitts came up with the model of an artificial neuron in 1943, and many computer scientists still use the basic features of this model.

There are normally five elements to an artificial neuron:

1. A set of inputs
2. A set of weights
3. A threshold
4. An activation function
5. A single output

These can all be numerical values, making it much easier to compute, and the output value is often expressed in binary form, with 1 corresponding to the firing of the cell, and 0 corresponding to no firing.

A Neuron Firing

So how can you create a functioning computer model of something as complex and little-understood as the human brain?

Firstly, neurons are nice to compute because they either fire, or they don't-the nature of an action potential means that there is no graded firing, so they can be represented in binary, and thus with on-off switches.

Secondly, current research is still focused on "lower" human functions, such as object recognition, for example, as there is still a lot to be learnt about these, so scientists have avoided some of the philosophical and computational issues associated with, for example, consciousness or emotion.

The real-world applications of neural networks can pretty much be split into two categories:
1. Cognitive Modeling and

2. Artificial Intelligence.

The objective of cognitive modeling is to try and understand the actual workings of the human brain by building computer models, and this is becoming especially useful as neuroscientists are finding that there is a limit to how much information we can gather from the stimulation and recording of actual live neurons, as the connections are so intricate etc.
 
So a neuroscientist may come up with a hypothesis for how a set of neurons would carry out a specific function in the brain, and a computer scientist would build a computer simulation to try and figure out whether this would work in practice. It could then be remodeled and refined depending on advances in research and the problems that come up.

Meera Desai, September 2008