J- +++++ ! ' ""'"'"' \"' """"' """''' \"' LLLLLLLLLLLLLLL j \"\"\"\"\"\"\" _._-_- """"""" """'' pairs to be learned must be presented in random order. Otherwise sequential dependencies may be built into the matrix. In fact, this problem rarely occurs. Some typical parameters might be useful to check program performance. Suppose we are working in a 200 dimensional system with perhaps 25 associations to learn. The system will learn essentially perfectly, that is, with the cosine of the angle between g and gT for every association over 0.9999 with about 25 presentations of each vector for a total of about 600 random presentations. The learning constant K was slightly less than l/fo. Since we present each pair multiple times, this simulation will obviously take many times the computer time of a system where each association is only learned once. However, if the simulation starts to take minutes something is wrong with the program. (Or you are using an extremely slow computer.)