The human brain contains approximately 100 billion neurons. Some of them connect to ten thousand other neurons. Together, they form neural networks (see picture). Each ‘unit’ depicts a neuron or a group of neurons. Usually, an artificial neural network is made up of three layers: An input layer, a hidden layer, and an output layer (Thagard, 2005). The input layer receives information, e.g. from our senses, and distributes the signal throughout the network, also known as ‘spreading activation’. The hidden layer does not have an initial interpretation, but serves an important role with respect to its connections with other units. The output unit passes information to other parts of the brain, e.g. to undertake the appropriate action in a particular situation. As an example, when we perceive an object, the input units receive certain properties like “brown, tail, four legs, long hair”. The output units will then able to classify the object as “dog”. Finally, the connections between units can have different strengths, called ‘weights’. These weights can either be positive, resulting in excitation of the neurons they connect to, or negative, resulting in inhibition. The mechanism of learning is, in essence, adjustment of the weights of connections (Thagard, 2005; McLeod, Plunkett & Rolls, 1998). How does a neural network represent knowledge of the world? There are two ways in which a connectionist model can store knowledge: Local and distributed. With local representations, each concept is encoded by a single unit (Plaut, 1999). This is not very likely, however, because it would imply the existence of ‘grandmother cells’, which states that one neuron would be associated with only one specific stimulus (LeVoi, 2005). A more realistic approach is the one of distributed representations, in which concepts are encoded by several units. Distributed representations of knowledge have a few advantages compared to local representations of knowledge. First, damage to a unit, by a head injury for example, does not lead to an immediate loss of all the knowledge stored in the network. This is known as ‘graceful degradation’. Because the concept is stored over several units, the network is still able to maintain the concept fairly accurate (LeVoi, 2005; Thagard, 2005). Second, local representations are economically efficient. That is, multiple concepts can be represented by only one neural network (LeVoi, 2005). The attractive qualities of connectionism as a cognitive theory are manifold. As for one, it is psychologically plausible. For instance, some models of connectionism accurately simulated human performance on word recognition tasks (Thagard, 2005). Furthermore, neural networks are capable of ‘content addressability’. Meaning that, just like humans, the network can bring up all the information that is needed, when it is presented with only a partial cue of that information (LeVoi, 2005). At last, neural networks can process more than one piece of information simultaneously. Therefore connectionism is sometimes referred to as ‘parallel distributed processing’ (PDP), in contrast to rule-based systems such as ACT-R, which operate in a serial fashion (LeVoi, 2005; Thagard, 2005). With the rise of the computer in the 1950s and 1960s, the view of the brain as a parallel information processor became very popular (Thagard, 2005). Later, the idea of neural networks appeared to be of major relevance in the development of artificial intelligence.
@ 2005
