neuron@tellab5.TELLABS.COM (Don Graft) (12/28/89)
I was wondering if anybody is aware of any work done on holographic paradigms in neural nets. Of course I am aware of work done in the late 60s and early 70s by such people as Westlake, Gabor, and Pribram, and of recent work implementing *optic* systems for heteroassociative recall and for (this amuses me) interconnecting traditional connectionist nets. However, noone seems ever to have actually simulated neural holographic heteroassociative memory and, as far as I can tell, nobody is even trying. Given the well-known shortcomings of back-propagation and the advantages of a holographic paradigm (truly distributed representation, invariance, high capacity) this surprises me. Why is this aspect of neural net research not benefiting from the recent explosion of interest in neural systems?????????? Donald Graft ...uunet!tellab5!neuron
bill@boulder.Colorado.EDU (12/29/89)
In article <1904@tellab5.TELLABS.COM> neuron@tellab5.TELLABS.COM (Don Graft) writes: >I was wondering if anybody is aware of any work done on holographic paradigms >in neural nets. > . . . >Why is this aspect of neural net research not benefiting from the recent >explosion of interest in neural systems?????????? > Well, Janet Metcalfe has been exploring a model she calls CHARM, which stores associations as the convolutions of the two inputs. I think that's how a hologram works, if I remember correctly. Unfortunately, I can't quote any references. Also, Kohonen's book "Self Organization and Associative Memory" contains a pretty extensive discussion of the relationship between neural networks and holography. -- Bill Skaggs
gary@desi.ucsd.edu (Gary Cottrell) (01/02/90)
In article <1904@tellab5.TELLABS.COM> neuron@tellab5.TELLABS.COM (Don Graft) writes: >I was wondering if anybody is aware of any work done on holographic paradigms >in neural nets. Of course I am aware of work done in the late 60s and early >70s by such people as Westlake, Gabor, and Pribram, and of recent work >implementing *optic* systems for heteroassociative recall and for (this >amuses me) interconnecting traditional connectionist nets. However, noone >seems ever to have actually simulated neural holographic heteroassociative >memory and, as far as I can tell, nobody is even trying. Given the well-known ... > >Donald Graft ...uunet!tellab5!neuron Janet Metcalfe has been modeling the hippocampal memory system as a discrete holographic system since 1981. The CHARM (Composite Holographic Associative Recall and Recognition Memory) model works by convolving pairs to be associated and then adding them into a memory trace. A single item is stored by convolving it with itself. Retrieval is done by correlation. CHARM accounts for a wide range of psychological data and successfully predicts new data. She has two Psych Review papers on it under the name Metcalfe Eich, in 1982 vol 89 p627-661, and 1985 Vol 92, p1-38, and a new paper on blended memories coming out in JEP General shortly. She had a poster at the recent NIPS conference that shows how her model accounts for "the function" in the recognition failure paradigm. The data in the recognition failure paradigm is so orderly that Tulving has called it a psychological law. "The function" (imaginative name!) is a function relating the probability that a test item in a paired associate memory task will be recognized given that it was recalled. You might imagine that if you recalled an item from a list, then you would recognize it, but this is not the case! There is only a slight dependence, and no mechanistic memory model has accounted for it except hers. I don't know if she has preprints to give out, but you can reach her at metcalfe@cogsci.ucsd.edu. The model is very interesting because storing an association is very fast. I think it has very good complementary properties to some of the abilities of a back prop learning system. gary cottrell 619-534-6640 Sec'y: 619-534-5288 FAX: 619-534-7029 Computer Science and Engineering C-014 UCSD, La Jolla, Ca. 92093 gary@cs.ucsd.edu (ARPA) {ucbvax,decvax,akgua,dcdwest}!sdcsvax!gary (USENET) gcottrell@ucsd.edu (BITNET)