In my D.Phil. thesis I am investigating the philosophical implications of the latest research in cognitive neuroscience and cognitive psychology, particularly connectionist modelling and dynamical systems theory. These areas are suggesting new ways of understanding the human mind, and most importantly, our conceptual capacities and the role of language. They are sensitive to the details of real neural processing, and thus can be considered to be 'neurally inspired'. This contrasts with - and is a direct reaction to - the more traditional field of artificial intelligence (AI) which explains cognitive operations in terms of syntactic mechanisms similar to those utilised by modern digital computers. The contrast is between an apporach that posits static structures and one that posits dynamic self-organising systems.
The main thrust of my research is concerned with refuting a set of arguments put forward by Jerry Fodor et al. which contend that a connectionist approach could never properly account for important aspects of linguistic thought. Specifically, it is claimed that connectionism cannot provide an adequate account of the systematicity of thought. This term is used to indicate the interconnectedness of linguistic capacities. For example, if a person can think the thought that 'John loves Mary', they can also, as a matter of brute empirical fact, think the thought that 'Mary loves John'. One simply does not find individuals who can only think one of these. Fodor takes this to show that thought must have compositional structure, which is something that he thinks computational systems have, and connectionist systems do not.
In my thesis I take on board Fodor's criticisms of current connectionist explanations. However, the central part of my thesis is the claim that connectionist models can be systematic if a new and more biologically plausible approach is taken. Part of this new approach is a novel kind of modularity - extending the work I did in my B.Phil. thesis on the increased computational power of multiple neural network models. Modularity is the idea that there are separate systems in the brain that have specific computational and functional roles. Fodor has argued that only input systems can be modular and that what he calls 'central processing' (for which read rational thought and other higher aspects of cognition) cannot be modular because of its special features. I think this is mistaken, and I argue that an altered notion of modularity, involving loose task specific coalitions of neural systems that constantly alter with context, can adequately explain 'central processing'. This approach can be modelled using many networks together, and not just single networks in isolation, as is currently the case in connectionist research. This interaction between multiple systems is exactly what is being found in the brain by cognitive neuroscientists. Brain imaging studies, for instance, reveal that even very simple linguistic tasks involve many brain areas, and which areas are active is dependent upon the exact details of the experimental task.
It is my hope that this research will lead to a new way of investigating and understanding the most interesting and essential aspects of human cognition, namely language and our ability to think about the world using it.