On modelling the retina using neural networks

The authors studied the retinal architecture and used it to build a system that would model and simulate the information processing that is being done in the retina. The architecture of the model is inspired by, and closely approximates, the actual neural architecture of the retina. The basic premise is that the neural architectures have evolved and fine tuned themselves for the particular tasks that they perform. Any attempt to duplicate them with an artificial neural system would then perform optimally if it were to use the same, or at least a similar, architecture. The model was trained using the backpropagation rule. The model was able to clearly replicate most of the important retinal functions. The simulations showed that, starting from small random initial values for weights, the network developed connections which were close analogs of the biological systems. The result of this development is an optimal or near optimal set of connections to perform the task