Realistic neural network models using backpropagation: panacea or oxymoron?

Abstract

Recent advances in neural network training algorithms make it possible to train network models of distributed processing systems of unprecedented complexity. After training, the response properties of interneurons in the model are often very similar to those in the biological system, suggesting that realistic network models are feasible. Prospects for further progress in using network training algorithms to construct realistic network models are examined by contrasting two examples of realistic network models trained by backpropagation. Recurrent backpropagation, which can train networks with feedback connections, holds great promise for realistic network modeling.