dario@techunix.BITNET (Dario Ringach) (12/08/89)
Almost all suggested motion detectors based on spatiotemporal energy measurements are sensitive to the contrast of the stimulus, so that they do not encode velocity information *directly*. Assuming that division is NOT a biological plausible operation, how can this problem be solved by means of a simple, biological plausible, neural network? Gradient methods suffer form the same problem in that division is needed to get rid of contrast dependency... Thanks in advance for any help! -- BITNET: dario@techunix | Domain: dario@techunix.technion.ac.il Dario Ringach, Technion, Israel Institute of Technology, Dept. of Electrical Engineering, Box 52, 32000 Haifa, Israel
bill@boulder.Colorado.EDU (12/09/89)
> Almost all suggested motion detectors based on spatiotemporal energy > measurements are sensitive to the contrast of the stimulus, so that > they do not encode velocity information *directly*. Assuming that > division is NOT a biological plausible operation, how can this problem > be solved by means of a simple, biological plausible, neural network? > Gradient methods suffer form the same problem in that division is needed > to get rid of contrast dependency... Thanks in advance for any help! > > Dario Ringach, Technion, Israel Institute of Technology, > Dept. of Electrical Engineering, Box 52, 32000 Haifa, Israel > But division IS a biologically plausible operation -- in fact it's actually more plausible than simple subtraction. The most common form of inhibition in the brain (GABA acting to open chloride channels) performs an operation more like a division than a subtraction (though it's really a sort of combination of the two). Neurophysiologists often refer to this as "shunting inhibition". Bill Skaggs
brp@sim.uucp (bruce raoul parnas) (12/09/89)
In article <14700@boulder.Colorado.EDU> bill@synapse.Colorado.EDU () writes: >performs an operation more like a division than a subtraction (though >it's really a sort of combination of the two). Neurophysiologists often >refer to this as "shunting inhibition". > > Bill Skaggs Actually, i believe that "shunting inhibition" is more akin to multiplication than either of the two operations you mention. the effect *is* to reduce the main path signal, as a division by a quantity greater than one would do, but the action is to divert "shunt" some of the signal through a multiplicative process. The shunting inhibition used by Grossberg in his contour enhancement and ART models takes the form: (B - xsubi)*f(xsubi) where B represents the total number of units, or populations, xsubi is the unit activity level and f(w) is a nonlinear feedback function from unit i to unit i. the effect of (B-xsubi) is to multiplicatively reduce the effect of the self- feedback when xsubi gets large to prevent saturation. bruce
ns299ad@sdcc6.ucsd.edu (Pablo R Alvarez) (12/10/89)
In article <8977@discus.technion.ac.il> dario%techunix.bitnet@jade.berkeley.edu (Dario Ringach) writes: >Assuming that >division is NOT a biological plausible operation, how can this problem >be solved by means of a simple, biological plausible, neural network? >Gradient methods suffer form the same problem in that division is needed >to get rid of contrast dependency... Thanks in advance for any help! In fact, depending upon the type of division you are talking about, division IS a biologically plausible operation. Consider the following: an inhibitory neuron receives inputs from n input cells and fires in proportion to the strength of that input. This inhibitory cell I contacts a neuron A. The output of A will be great when the input to I is small, and vice-versa: this system performs a division operation. Note: in fact, you'd need a lot of inhibitory interneurons to really do this job right, and it isn't, of course, an exact division. However, it could very well do the job. There are similar circuits in the brain, and it has been hypothesized that this might be their function (McNaughton and Morris, Trends in Neuroscience, 1987 or 88, I can't remember right now). Pablo Alvarez (palvarez@ucsd.edu)