INS_ATGE@JHUVMS.BITNET (04/13/88)
Something which has recently struck me is how little programming is actually done in neural networks. An experiment was recently done by Dr. Sejnowski of JHU regarding the interneurons or "hidden units" which develop in a neural network which is trained to recognize concave vs. convex features on sight. Now although I'm sure the researchers had some guesses as to the "receptive" and "projective" areas of the hidden units, they never "programmed" them. The neural network was trained (using back propogation or possibly a local minima avoiding algorithm), and the hidden units ended up looking like neurons found in the cat visual pathway which correspond to a subclass of what were originally thought to be edge-detection neurons (note: the concave vs. convex neruons are a subset of the so called "complex" cells thought to be edge detectors--not all the data on complex cells fits all the concave vs. convex hidden units, so it appears they are a subclass (particularly, the subclass which do not respond well in the center of field)). In other words, although the experimenters tried to create something, they did not "program" the entire system--it organized -itself-. Not only that, but it added a new hypothesis to the area of neuroscience. Thus this new area of science is labeled "Computational Neuroscience." -Thomas Edwards