Towards a Non-Network Approach to Neural Modeling: Some Basic Issues of Measurement, Simulation and Computational Significance of Brain Maps

摘要

The term neural network is often associated with the construction of networks of schematic neurons to implement functions such as associative memory, classification, visual segmentation, etc. One advantage of this approach to modeling the nervous system is its explicit computational power: neural models are set up with the exclusive goal of solving a particular computational problem. One disadvantage of this approach is its remoteness from the actual data of the nervous system. The details of complex neural networks are very difficult to observe. The experimental likelihood of such observation, or even of experimental constraint on current network models, is not favorable in the near future. This work provides a non-network approach to neural modeling in the following sense: we model brain architecture and computation at a continuum, rather than a discrete, or neuronal, level of scale. From a practical point of view, this allows us to simulate biological processing of early vision using conventional image processing techniques (e.g. convolution), avoiding the largely unknown details of network level implementation. In doing so, a concept of the cortical component of the brain as a map machine emerges. That is, novel architectures which have been observed over the past few decades may have computational significance, and may represent an alternative approach to the theory of neural computation which stresses the aspect of data structure over the details of network implementation. (KR)