Comparison of brain structure to a backpropagation-learned-structure
Mario A. Serna, Leemon C. Baird · 2002
This paper describes the results of experiments studying the circumstances under which an error-minimizing artificial neural network mimics the mammal visual system. The networks were trained to recognize handwritten-digits. The experiment was not intended to yield a high identification accuracy, but rather to generate a comparison of the neural networks to biology under different circumstances. Experiments were conducted with partially hand-set networks, freely-trained networks, and convolutionally constrained networks. The convolutional experiment, where certain weights were constrained to be identical, performed the best at digit recognition while also modeling parts of biology that we had not anticipated the network would model. Rather than using the input image to generate an edge detection outline, as occurs in the retina, the network's first layer modeled the cones themselves, reacting most to one color (black or white), but not performing any real processing.