It is the name for the computer modelling approach to information processing based on the design or architecture of the brain. Connectionist computer models are based on how computation occurs in neural networks where neutrons represent the basic information processing structures in the brain.
All connectionist models consist of four parts:
- units: they are what neutrons are to the biological neural network, the basic information processing structures. Most connectionist models are computer simulations run on digital computers. Units in such models are virtual objects and are usually represented by circles. A unit receives input, it computes an output signal and then it sends the output to other units. This is called activation value. The purpose of the unit is to compute an output activation.
- connections: connectionist models are organised in layers of units, usually three (3). A network however, is not simply an interconnected group of objects but an interconnected group of objects that exchange information with one another. Network connections are conduits. The conduits through which information flows from one member of the network to the next are called synapses or connections and are represented with lines. (in biology synapses are the gaps between neutrons, the fluid-filled space through which chemical messengers -neurotransmitters- leave one neutron and enter another)
- activations: activation value in connectionist models are analogous to a neuron’s firing rate or how actively it is sending signals to other neurons. There is a big variability between the least active and the most active neutrons expressed in a scale fro 0 to 1
- connection weights: The input activations to a unit are not the only values it needs to know before it can compute its output activation. It also needs to know how strongly or weakly an input activation should affect its behaviour. The strength or weakness of a connection is measured by a connection weight. They range between -1 to 1. Inhibitory connection reduce a neuron’s level of activity; excitatory connections increase it.
Yet, the behaviour of a unit is never determined by an input signal sent via a single connection, however strong or weak that connection might be. It depends on its combined input. That is the sum of each input activation multiplied by its connection weight. The output activation of a unit represents how active it is, not the strength of its signal.
Connectionist networks consist of units and connections between units and have some very interesting features like emergence of behaviour. This does not reduce to any particular unit (liquidity in water). Graceful Degradation and Pattern Completion are two ways in which activations are spread through a network. They are not classical computers, their behaviour does not arise from an algorithm, they learn to behave the way they do.
Robert Stufflebeam, 2006. Connectionism: An Introduction (pages 1-3), in CCSI (Consortium on Cognitive Science Instruction) supported by the Mind Project, full article available here
Image available here