jagota@sybil.cs.Buffalo.EDU (Arun Jagota) (05/25/91)
For associative memories, sizes of basins of attractions has been an object of theoretical study, motivated by the fact that large basins are ``better'', for error-correction abiility. While not disputing this, I offer one (error-detection) application for which basin attraction sizes is not an issue, whereas the ``stability'' of the stored patterns is. In one experiment, we ``stably'' stored ~240,000 English words in a ~580 unit fully connected Hopfield-type Network. The test set to the network was a collection of random non-word strings. The task was to identify the strings as errors. The network identified between 80-90% of the test set strings as errors, taking ONE network cycle per string. A plausible application of this idea is to take a large document containing words from a LARGE dictionary, and ``garbage'' strings (far from any dictionary word) and rapidly remove (many of the) latter. shar file of LateX sources of the paper describing the details is available via ftp as folows. ftp ftp.cs.buffalo.edu -- if this doesn't work try the following number, which will change -- in a few weeks. Or send me e-mail. ftp 128.205.32.3 Name : anonymous > cd users/jagota > get ijcnn_Jul91b.shar > quit There are several related papers in the same directory, accessible by ftp as above. Get the file `README' for details. The simulator is also available there. Get the file `hsn.README' for details. Finally, as a last resort, but please, only if all else fails, or if you do not have access to ftp, send me e-mail (jagota@cs.buffalo.edu) and I can send you the same files (papers or simulator) by e-mail. -- ------------------------------------------------------------------------ Arun Jagota Internet: jagota@cs.buffalo.edu Computer Science Department BITNET : jagota%cs.buffalo.edu@ubvm.bitnet State Univesity of New York at Buffalo