eliot@phoenix.Princeton.EDU (Eliot Handelman) (11/01/88)
>Subject: Music and PDP (II) >>From: MUSICO%BGERUG51.BITNET@CUNYVM.CUNY.EDU >Date: Wed, 19 Oct 88 15:23:00 +0100 > >There are other papers on Music and PDP. See : > J. Bharucha : "Neural Net Modeling of Music", > M. Leman : "Sequential (Musical) Information Processing > with PDP-Networks", > B. Vercoe : "Hearing Polyphonic Music with the Connection Machine" >in : Proceedings of the first workshop on Music and AI. > >Work is also done by C. Lischka (See the Proceedings of the ICMC 87 and the >Arbeidspapiere der GMD). I have two other related papers ("Neural Net- >works in Music Research" and "Massive Parallel Computer Methods in Music >Research"). I currently revise these papers and they will very soon be >available. > >I guess that many other people are working in the same direction. > >Marc Leman >University of Ghent >Institute for Psychoacoustics and Electronic Music >Blandijnberg 2 >B-9000 GHENT >Belgium I happened to have recently studied the three papers that Mr Leman mentions. I have some remarks to make about both Mr Leman's and Mr Bharucha's papers. To begin with, Mr. Leman notes, in "Sequential (Musical) Information Processing with PDP-Networks," that there has been some recent interest in the "possibility of processing sequential information." and promises to show us how to "store and process" "musical concepts," which he calls "constraint examples of time-varying information." Fair enough. What is unclear to me is the nature of Mr Leman's research. Is he proposing to adapt other research to his specific needs, or is he merely providing us with a summary of extant research? Mr Leman writes, for example, that the "sequential autoassociator [..] is similar to the class of pattern associators described by [...] Kosko (1988) BUT [my emphasis] applied to the temporal domain." Is the suggestion not clear that this application is original? I was curious to see what sources Mr Leman was drawing upon, and so I looked up some of his references. One of these was a paper by Bart Kosko in the IEEE Transactions on Systems, Man and Cybernetics, Vol. 18 no. 1, Jan/Feb 1988, "Bidirectional Associative Memories," which appeared in print just a few months before Mr. Leman's paper. Kosko: "The natural suggestion then is to *memorize* the association (Ai,Bi) by forming the correlation matrix or vector outer product AT/iBi. [...] The next suggestion is to *superimpose* the m associations by simply adding up the correlation matrix pointwise: [a formula is given]." [Kosko,52] Leman: "Associations can be memorized by taking the vector outer product of the associated patterns and then superimpose them so that: [the same formula is given]." It seems quite clear that Mr Leman has read Mr Kosko's paper. Did he not notice that Mr Kosko is specifically interested in applying his asscociator to the temporal domain? In his abstract, Mr. Kosko writes: "Temporal patterns are represented as ordered lists of binary/bipolar vectors and stored in a temporal associative memory (TAM) ..." [Kosko, pg 49]. Just in case there was any doubt that Mr Kosko is interested in music, he writes that "...a sequence of binary vectors can represent a harmonized melody." [pg 49] Not fair, Mr. Leman. Credit where credit is due. Mr Bharucha's paper is another matter. Mr Leman at least gave us a few formulae; Mr Bharucha is content to inform us of the power of his various architectures and is apparently unwilling to let us judge for ourselves. He writes, for example: "This simple architecture has extraordinary predictive power for harmonic expectancies in Western music. It predicts the build up of a tonal context over the course of a sequence and it preserves the functional ambiguity of individual events, thereby supporting smooth modulations." These claims are nothing less than extraordinary. What exactly is meant by "Western music"? Would this include the Trauerzug of Act II of Parsifal, for example? How does it predict the build up of a tonal context? How does it "preserve the functional ambiguity of individual events"? In what way does it "support smooth modulations"? Mr Bharucha feels that it is unnecessary to substantiate any of these claims with a few examples; I feel that it is. Is Mr Bharucha providing us with new architectures? This seems doubtful. In "Storage and Processing of Information in Distributed Associative Memory Systems" by Kohonen et al., in "parallel Models of Associative Memory," Hinton & Anderson, eds. LEA, 1981, page 121, an architecture is discussed that recalls temporal sequences with additional context inputs using delayed input that closely resembles those proposed by Mr. Bharucha. If Mr. Bharucha is claiming to do original research into the architecture of temporal pattern associators, he is at least 7 years out of date; if he is doing some sort of music research, he might credit us with some curiosity as to the basis of his claims. Eliot Handelman Department of Music Princeton University Princeton, NJ 08540