Subject: Re: 40 Hz RIP From: DeLiang Wang <dwang(at)CIS.OHIO-STATE.EDU> Date: Sun, 25 May 1997 16:13:09 -0400Hi Neal, > This is also the basis of the difference I have had with the so > called "oscillatory framework" which it seems to me is a case > of putting the oscillator cart before the signal horse. The > information for binding is already there is the common temporal > structure of the inputs from receptive fields which are > activated by a common source. Having a whole load of > "oscillators" on top of this is, quite frankly, massively > computationally redundent. I simply cannot believe that natural > selection would come up with such a scheme. I wish you understood the oscillatory framework before you attacked. The information for binding is, of course, in the input and memory. The whole point of binding is HOW to extract and represent the information scattered around in neural signals. With or without oscillations, the binding problem is real if one thinks of distributed representations. There is no mystery to oscillations, particularly from the modeling perspective. Neurons generate spikes in response to stimulation. The Hodgkin -Huxley equations (published in 1952) and their later simplifications all lead to periodic firing of action potentials, i.e. oscillations, with constant input. It is in this sense that we talk about oscillations for modeling (as a mathematical concept), and it does not mean periodic activity all the time (e.g. changing stimuli lead to non-periodic firing). Indeed our oscillator models are derived from the Hodgkin-Huxley equations. Well, natural selection seems to come up with highly nonlinear neuronal spikes. Any computations with such spikes would need to incorporate some aspects of oscillations. > > One important thing which many people seem to forget about the > visual system is that if an image is actually stabilised > ...... It's better to leave the above speculations to the vision community. > > Those who are advocates of the "oscillatory framework" have > searched in desperation for some evidence of "40 Hz > oscillations" in the auditory system, but to no avail. How many have looked in the auditory system? > > As you say Peter, if you talk to vision people about this they > are becoming increasingly sceptical about the Singer et al > results, at least in part because it has been difficult to > replicate. Perhaps now the 40 Hz oscillator business can be > finally put out of its misery, but somehow I think it may > require some more anaesthesia before it may finally rest in > peace. I heard such predictions by prominent neurophysiologists in 1991-1992 saying that neural oscillations would die in 1-2 years. Things have not turned out that way, to the disappointment of some. Evidence in awake animals and in other animal preparations has been reported. Here are several recent papers that have found coherent oscillations: C. Gray & D. McCormick, Science, 274, 109-113, 1996; M. Livingstone, J. Neurosci., 75, 2467-2485, 1996; P. Roelfsema et al., Nature, 385, 157-161, 1997). The article by Margaret Livingstone, in particular, contains explanations why some studies haven't seen oscillations. For a hypothesis as fundamental as temporal correlation, it is impossible to draw either a positive conclusion or a negative conclusion from several studies. Those who wish to see it gone quickly may be disappointed again. Think about the old and ongoing debate between local representations and distributed representations - still unsettled. Whether temporal correlation turns out to true, partially true, or entirely wrong, exploration of oscillatory dynamics to solving the binding problem is worthwhile. As I said in my previous message of May 20, it is theoretically an elegant way of representing multiple streams and it is a new style of computation. But computing grouping and segregation is still a big challenge even if temporal correlation as a representation is true. Again think about local/distributed representations. Whether or not either or both are true, it will not give us computational answers automatically. Otherwise wewould only need three computational people (one for local rep., one for distributed rep., and the third for the combined rep.), and the field of neural networks would have accomplished its mission long ago. So given tremendous computational tasks ahead, let's keep alternatives alive before the problem is solved. If this debate is to continue, it would be more productive to focus on technical aspects instead of trashing any approach based on personal sentiments. Regards, DeLiang Wang