Neural Computation and Symbolic Thought The Philosophical Implications of Biological Plausibility in Connectionist Network Modelling

features of its input. This structure in the hidden layer was then utilized by the output layer to produce the appropriate responses as defined by the assigned task. Considering Chomskian arguments about the poverty of the linguistic stimulus available to infants, this network should act as a warning that the learning environment may in actual fact carry much more information than was previously thought. Distributed representation not only gives rise to complex processing abilities, it also means that networks are operationally robust. As representation and computation involve many units and connections, each individual unit or connection plays only a small part in the overall calculation. As a result, a network can afford to lose a few connections or units whilst still attaining a reasonable level of performance. It can also cope with incomplete or noisy inputs. The performance gradually declines with the number of elements that have been removed, or the disturbance to the input, hence the processing of distributed networks is described as displaying ‘graceful degradation’. This feature is noteworthy because of its relevance CONNECTIONISM AND NEURAL COMPUTATION 17 to biological systems. First real neurons are inherently ‘noisy’ in that they have a basal spiking frequency from which they deviate according to a normal distribution. Thus a neuron may fire more vigorously, in a way that would normally have some significance for its receiving neurons, even when none of its preferred stimuli are present. Second, it is an unfortunate fact of life that brains lose neurons throughout life, whether due to injury or the natural course of ageing, and so must be able to continue to function in the face of such loses. Put bluntly, the connectionist argues that, as brains degrade gracefully, anything claiming to model the brain had better do the same. The contrast here is supposed to be with conventional computers, which will crash even if only one line of a program, or one transistor, is missing or broken. While this is a persuasive argument it is far from conclusive, I merely mention it as an important debate between parallel distributed processing and conventional computational models. 1.2 Vector Cognition A bare description of the functioning and abilities of real and artificial networks does not count as philosophy. To be relevant to philosophy the processing of these networks must be related to mental processes occurring at the personal level. Paul Churchland has suggested that cognition should be modelled on the way basic perceptual processing works. According to his view of such processing, sensory inputs are mapped onto the appropriate prototype, which constitutes the creature’s understanding of that input. Thus the bridging principle between neurocomputation and the philosophy of mind is that of the prototype. In this section I will give a brief sketch of Churchland’s approach before going on to show that it, and other standard connectionist models, cannot fully account for human cognition in section 1.3.