luong@mirsa.inria.fr (Tuan Luong) (04/09/90)
Consider a neural network with several local minima of the energy function, each of which representing a different pattern, and having differents slopes and depths. Consider now an image containing different patterns among those represented in that network. Each of these ones will be recognized but the different local minima will be reached at different times. This means that vision in a temporal process: object 1 is perceived, then object 2 ... So, how to create a representation of a composed object made of an assembly of different patterns ? One can say that the temporal synchrony of the units is a representation of the global identity, but how to obtain this synchrony between the different parts of the image ?
wipke@secs.ucsc.edu (W. Todd Wipke) (04/10/90)
>Consider a neural network with several local minima of the energy function, >each of which representing a different pattern, and having differents slopes >and depths. >Consider now an image containing different patterns among those represented >in that network. Each of these ones will be recognized but the different >local minima will be reached at different times. This means that vision in a >temporal process: object 1 is perceived, then object 2 ... >So, how to create a representation of a composed object made of an assembly >of different patterns ? >One can say that the temporal synchrony of the units is a representation of >the global identity, but how to obtain this synchrony between the different >parts of the image ? Chemists routinely minimize energy of three-dimensional networks which can have several local minima. Each minimum represents a low energy state of the molecule, in effect, the "personality" of the molecule. I wonder if there are any interesting conclusions one can draw. On another but related topic, chemists have successfully used learning machines for chemical pattern recognition--the work was done in 1969-71. Very systematic studies of learning rate versus feature scaling, number of parameters, type of feedback, size and diversity of training set etc. Peter Jurs, Penn State wrote a book on it and many papers showing one can predict mass spec, determine molecular formula from mass spec, classify drugs, etc. From the literature I have seen, this work has gone unnoticed by the computer science community. Since the data sets are well defined, it would provide a reproducible standard against which you could all compare your methods or black boxes. To my knowledge there is no such standard in use today. I would be very interested to see if today's methods are better than earlier ones. I have not seen systematic studies like Jurs did, but would like to see some. ======================================================================= W. Todd Wipke wipke@secs.ucsc.edu Molecular Engineering Laboratory wipke@ucscd.ucsc.edu Thimann Laboratories wipke@ucscd.bitnet University of California BBS 408 429-8019 Santa Cruz, CA 95064 FAX 408 459-4716 ======================================================================= ======================================================================= W. Todd Wipke wipke@secs.ucsc.edu Molecular Engineering Laboratory wipke@ucscd.ucsc.edu Thimann Laboratories wipke@ucscd.bitnet University of California BBS 408 429-8019 Santa Cruz, CA 95064 FAX 408 459-4716 =======================================================================
park@usceast.UUCP (Kihong Park) (04/12/90)
Why make the assumption that segmentable "parts" of an image are stored as different local minima on the same network, i.e., its energy landscape? You may want view your image processing system as consisting of a number of relatively "independent" modules, each of which with different functionalities. Then, temporal synchronicity can possibly be achieved. That is, in the most simplest case where each model encodes among other things one "part" of the image, the simultaneous convergence to local minima in each module may bring forth a synchronized convergence of the total system to a global minimum. Hence no temporal sequencing. The above explanation is of course too simplistic, but nevertheless it should illustrate that in any nontrivial system, modularization is a key factor. How to achieve this is a big problem. Kihong Park. (park@cs.scarolina.edu)