Subject: Re: 40 Hz RIP From: Neil Todd <todd(at)HERA.PSY.MAN.AC.UK> Date: Mon, 26 May 1997 21:04:43 +0100Hi DeLiang > I wish you understood the oscillatory framework before you attacked. In fact I have been following this for some time. I have a copy of your paper in Cognitive Science (I refer to it in Todd (1996) "An auditory cortical theory of primitive auditory grouping". Network:Computation in neural systems. 7, 349-356. ) and Guy Brown's various papers. As you know I have worked with Guy myself so I am quite familiar with the arguments. We have been arguing about this for two years now. You may also like to know that because I am so interested I actually teach this stuff to our undergraduates and postgrads. We have a second year course "Recent Papers in Perception" which includes Singer and Gray (1995) "Visual Feature Integration and the Temporal Correlation Hypothesis". Ann. Rev. Neurosci. 18, 555 and a MSc Cog. Sci. module "Machine Perception" which includes an analysis of four different models of auditory grouping (Beauvois and Meddis, 1991; Brown and Cooke, 1995; Todd, 1996; McCabe, 1997). > 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. Exactly! So, far we are in complete agreement. > 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) ..... It doesn't matter what kind of fancy oscillators you have, my point was how you arrange them in relation to the signal. For example, in Guy's chaotic oscillator the output of the oscillator had no temporal relationship with the signal whatsover. This is in direct contrast to what is known about cortical cells which phase-lock , depending on their modulation transfer function and the period of the stimulus. (Perhaps Guy might like to comment here? Come on, join the fun!) > 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. Yes, but I don't know of any natural system whose output bore no relation to its input! If you had looked more carefully at my message, particularly at the end I said that the impulse response of a modulation filter is a damped oscillation. The essential difference between my model and your approach is that in my case the function of the filters is to passively transform the signal into the frequency domain, i.e. to faithfully represent the temporal information in the signal *spatially* in terms of its power spectrum, including phase, rather than some more abstract relationship such as a clocking cycle. In order to do this you need a population of cells/filters which sample the range of periodicities that are found in natural signals. In the case of speech there are two such important ranges (a) voice pitch (b) voice rhythm. These ranges happen to coincide with the known ranges of BMFs in two important physiological stations in the auditory system, the ICC and the cortex (Langner, 1992. "Periodicity coding in the auditory system". Hearing Research, 60, 115-142.). > > > > 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. In fact Deliang, the information about the various kinds of eye movements required to avoid image adaptation was taken directly from Barlow and Mollon (1982) "The senses". CUP. The chapter was "Eye movements and strabismus" by Mathew Alpern. The information concerning the temporal and spatial contrast sensitivity functions was taken from the chapter by Woodhouse and Barlow "Spatial and temporal resolution and analysis". If I may quote from Hawken et al (1996, "Temporal-frequency selectivity in monkey visual cortex." Visual Neuroscience, 13, 477-492.) p. 490. "Human and monkey observers can perceive flicker and drift rates greater than 40 Hz under photopic conditions. Recent experiments indicate macaque retinal ganglion cells show flicker responses at higher temporal frequencies than the psychophysical critical fusion frequency (CFF). Under the assumption that CFF is dependent on cortical processing, then there must be additional filtering in the cortex because signals present in the retina and LGN are not available perceptually. Therefore, it seems likely that CFF depends upon cortical as well as subcortical temporal integration. The temporal tuning of a number of V1 cells would give a reliable response at the highest temporal frequencies that are required to account for human and monkey perception". It follows logically from this that a number of V1 cells will respond to micronystagmus. > > > > 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? Certainly, Gray and Singer have! In Ann. Rev. Neurosci. 18, p. 567 there is a section entitled "Evidence for synchrony in non-visual structures". There are the usual references to the hippocampus and the olfactory system. In the case of the auditory system there are four references to MEG and EEG evoked potential studies which they claim show evidence for gamma-frequency components. > 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, .... As I said, no doubt it may live on a little while yet, for some people. > For a hypothesis as fundamental as temporal correlation, it > is impossible to draw either a positive conclusion > or a negative conclusion om several > studies. ... As I also said, in fact I have no problem with the notion of temporal correlation. The disgreement I have is in relation the following three points: (a) what information is being correlated (b) how is that information represented and (c) by what mechanism is the correlation carried out? The solution I have proposed is (a) it is the temporal information of the signal, rather than abstract properties of a fancy oscillator (chaotic, van der pol, etc, etc) (b) the temporal information is represented *spatially* in the frequency domain ( a wavelet transform is one way to do this, but note that the impulse response of the basis functions are a damped oscillation, e.g. Gabor wavelet) (c) the correlation is carried out by a network of cortico-cortico connections where a basic processing unit is a cortical column As I also said, the general architecture of this model is not dissimilar to some "neural oscillator" models, including your own, but there is a fundamental difference in that it is based on the strong evidence that the brain represents its inputs in the form of multi-scale decompositions. To give you another concrete example of this let us return to the visual system. No one would now dispute that the visual cortex does a spatial-frequency analysis. These may be represented by a family of Gabor wavelets. The evidence seems now extremely clear that the visual cortex also does a temporal-frequency analysis (e.g. Hawken et al, 1996, as above). Joint spatio-temporal analysis provides the basis for visual motion perception. One model which uses exactly such a basis is that proposed by Heeger (1987, "Model for the extraction of image flow." J. Opt. Soc. Am. 4(8), 1455.). He used a family of three-dimensional (space-time) Gabor filters which effectively sample the power spectrum of a moving texture. The model has four stages (a) center-surround filter (b) 3-D Gabor filter (c) motion energies (d) velocity tuned units This model is entirely consistent with the physiology (a) bipolar cells (retina) (b) simple cells (V1) (c) complex cells (V1) (d) MT units (V4, medial temporal lobe in the primate). In case you hadn't already noticed from my first message, this provides the basis for a model of audio-visual binding since the temporal information for motion is coded in precisely the same form as for rhythm. Well I guess I've let the cat out of the bag. So I'd better rush this one into publication before I get trampled by the stampede. I've burnt my fingers with patents once too often to be bothered about prior disclosure. > So given tremendous computational tasks ahead, let's keep > alternatives alive before the problem is solved. DeLiang, I would be extremely upset if you gave up on "neural oscillators" - it would ruin my fun. > If this debate is to continue, it would > be more productive to focus on technical aspects I think you will realise that I have done extactly that. Best wishes Neil