Classical and Connectionist Cognitive Models
By ai-depot | September 2, 2002
Connectionism
Connectionism
Connectionist models are structured on the concept of neural networks. Such a network consists of patterns of units, or nodes, that are linked together by numerous pathways. They are called neural networks because they are analogous to the synaptic structures within the brain. There are generally three classes of nodes: input nodes, hidden nodes and output nodes, as shown below:
Figure 1: A neural network
Input nodes accept information from the world, output nodes show the result of processing, and the hidden nodes in between carry out the processing. Nodes receive signals from other nodes in the network and output signals to other nodes based on an activation function. This function uses two values: the weights of the connections between the nodes sending the signals and the node that is receiving them; and the threshold value of the receiving node. Depending on the function, the receiving node may fire off a signal if the threshold is exceeded by the sum of the weights of the incoming signals, or send different strengths of signal depending on how “activated” the node is.
When a network is in operation, its input nodes will be activated via some mechanism, such as a response to the pixels in a photograph, which will in turn activate the hidden nodes depending on the topography of the network. An output will be generated on the output nodes as a result of the propagation of signals throughout the network.
There are two views of how nodes form representations. The localist view is that a single node represents a single idea. The distributed view is that representations are spread over a number of nodes. In the latter, which is commonly accepted as a more useful method, groups of nodes would therefore represent ideas. Figure 2 below shows how this might occur in a model of reading.
Figure 2: An interactive-activation model of reading with distributed representations for words (from Lormand, 1991, originally adapted from Rumelhart and McClelland, 1982). In a localist network, words would be represented by a single node.
The network in Figure 2 cognises that the word “RUN” is present when a group of nodes is activated in a particular pattern. As a model of human cognition, this can be seen as a representation of an attitude about the word “RUN”, in this case the belief that the word is present. Other attitudes about the word would be represented by similar groupings, such that they might include some of the same nodes. These groups would essentially be overlapping, in the same way that concepts overlap. Overlapping provides an intuitively plausible explanation of pattern recognition, experience acquisition and knowledge distribution.
Connectionist networks also provide an account for the expert phenomenon in two ways. The first is in distributed parallel processing (PDP): the brain is a massively parallel computational device that makes up for slow neurons by firing off millions of signals at once. The second and more fundamental way is by removing the need for syntactic complexity as required by the LOT hypothesis. If representations are simple, the time required to perform inference over a number of propositions is considerably less than it would be for a system of complex representations. In a similar vein, one of the central tenets of symbolic systems is the concept of atomic and molecular representations. The difficulty with atomic representations is understanding how the meanings of these individual concepts come to be fixed in the first place. Connectionist networks essentially circumvent this concern due to their distributed nature. There is no way to distinguish between simple and complex representations in a distributed network when all such representations appear as patterns of nodes. In this sense, connectionist networks are considered to be sub-symbolic.
Tags: none
Category: essay |
