I had written:
> But what if the stimulus-driven
> synchronies are what the system uses to encode edges and
> spatial intervals between edges? One has thrown out the most
> important aspects of the neural response.
Jim Kroger wrote:
> I think it depends on what you want to include as
> a representation of the stimulus. One could argue
> that there are neural responses to stimuli which
> do not contribute to the representation of what
> matters about the stimulus to the perceiver, and
> one could argue that all response is part of
> the representation, but we selectively attend
> to part of it. If we do not have a way to eliminate
> that which we choose to call noise, we are unable
> to learn how the part of the neural response that
> we care about is represented apart from that noise.
I wholeheartedly agree with all of this. The whole question of stimulus coding
and neural representations concerns which aspects of the neural activity
subserve which perceptual qualities and operations on them. The neural
representations consist of those aspects of the response that are relevant to
a given function. I adopt the first view, that there are always some aspects
of the neural response that are not relevant to any one given perceptual
function (but they may subserve some other function). In order to identify
which aspects of the neural response are functionally-relevant I think the
most direct way is to examine how different aspects of the neural response
covary with what is perceived. For our work on pitch, we use stimuli with
different power spectra that evoke the same low pitch (i.e. they are metameric
stimuli; one could do this for light spectra that produce similar colors or
for similar visual shapes (triangles) formed from different elements, such as
dots, lines, color/texture contrasts, etc). In the case of pitch the aspect of
the neural response at the level of the auditory nerve that covaries with
pitch appears to be which all-order interspike interval is most common across
the whole population. (Cariani & Delgutte, Neural correlates of the pitch of
complex tones. I & II, J Neurophysiology, Sept 1996).
I (Cariani >>) had written:
> >I have never understood why synchronies that are not related
> >to the stimulus are so highly prized, but those related to the
> >stimulus are routinely discarded. Could you shed some light on
> >this?
Jim Bower (>) replied:
> A complete answer would require a long response, however, the short (flip)
> answer is that several theorists early on proposed spike synchrony as a
> means of solving "the binding problem". Then, when two research groups in
> German showed evoked potential synchrony under suggestive (of binding)
> stimulus conditions, we were off and running. It was a short step from
> there to the suggestion of several other theorists/philosophers that they
> could go whole hog and also solve the problem of visual attention, and (why
> not) consciousness at the same time.
I'd be very interested in the long response sometime. I myself think that the
re-introduction into neuroscience of synchrony, periodicity, and problems of
perceptual organization was a very welcome development, a breath of fresh air.
What I haven't understood is why the stimulus-locked cross-correlations aren't
considered as a way of representing spatial structure.
> Obviously, synchrony induced by the stimulus alone is not sexy enough to be
> a basis for such higher mental states, which (by intuition) must somehow be
> separate from pure stimulus response (we are free, aren't we?).
What's sexy is in the eye of the beholder of course, and there's no accounting
for taste, but what would be the barrier to basing the low-level
representation of spatial configuration (of relations between edges) on the
temporally-coherent volley pattern that is created when the edges of an object
cross the receptive fields of retinal elements at their respective relative
positions? Maybe the objection has been that the response latencies of
neighboring retinal elements are varied, so that one does not get strict
synchrony, but nevertheless, these kinds of latency differences could be
offset by all sorts of delays further in the pathway. In any case, the retinal
correlate of a contrast gradient would be the presence of
temporally-correlated spikes in a local retinotopic neighborhood. When there
is uniform luminance, there is uncorrelated or much less correlated activity.
If the visual object consisting of many edges is rigid, then the boundaries of
the object will produce correlated activity in the respective locations of all
the boundaries.
A low-level representation based on stimulus-driven temporal correlations
(zero or nonzero delay) doesn't mean that the all of the rest of the
processing must slavishly follow the stimulus, so our freedom is preserved.
> A few spurious, opinionated, and no doubt controversial and annoying remarks:
One of the most basic qualities of life is irritability, as Howard Pattee is
fond of saying......
> The 'binding problem' is a problem in feedforward nets that comes up when
> you try to parse visual signals into objects (i.e. in machine vision) --
> the brain is not feedforeward, and therefore I doubt feature binding is a
> serious problem at all. (In model-based systems, things are bound to begin
> with).
I do think that recurrent networks are essential for perceptual
organization. I have been working on timing nets that build up
recurrent temporal patterns in their inputs (and separate multiple
patterns that have different fundamental frequencies). These all
group by common, recurring phase structure, which seems to me
to be what one needs for image stabilization and object formation
in vision. When different visual objects (two transparencies) move
relative to each other, each object has its own set of invariant
spatial relations between its edges that recur again and again
in different visual locations. These patterns would get built
up, while patterns between edges in different objects would be
constantly changing so they would not build up (fuse).
> Several authors have pointed out in the past that one has to be careful
> interpreting cross correlations. In our own work published with Matt
> Wilson ten years ago, we concluded based on a realistic network simulation
> constructed to examine the mechanisms likely to underlie the data obtained
> in Germany:
>
> "Interpreting phase coherence from correlation functions produced from the
> average of many simulation trials pointed out the need to distinguish
> average phase effects from instantaneous phase effects. Instantaneous
> phase implies that the statistics of the correlation function taken at any
> trial are consistent with the statistics of the combined data. Average
> phase allows for systematic within-trial and between-trial variability and
> is, therefore, a weaker assertion of actual coherence. This distinction is
> particularly important for theories which rely on phase encoding of
> stimulus information. Analysis of our model results indicates that the
> observed phase relationships are an average, rather than an instantaneous
> effect."
>
> Two related papers (quote from the first):
>
> Wilson, M.A. and Bower, J.M. 1990 Computer simulation of oscillatory
> behavior in cerebral cortical networks. In: Advances in Neural information
> processing systems. Vol. 2, D. Touretzky, editor. Morgan Kaufmann, San
> Mateo, CA., pp. 84-91.
>
> Wilson, M.A. and Bower, J.M. 1991 A computer simulation of oscillatory
> behavior in primary visual cerebral cortex. Neural Computation 3: 498-509.
One has to be careful interpreting cross-correlations.
I take it that these papers deal with the question of phase coherence
where one is not taking into account stimulus-driven coherence. Were
the inputs in the simulation temporally structured?
> With respect to stimulus or nonstimulus driven synchrony, I suggest we
> consider the somewhat larger question, why, in principle, are synchronous
> spikes more important than asynchronous ones to begin with. In other words,
> why don't all the spikes code information? The common answer is that two
> action potentials arriving at the same time on a postsynaptic cell have an
> increased probability of inducing a spike in that neuron. However, this
> assumes that neurons are simple integrate and fire devices (with no
> dendrites), which ample evidence suggests they are not. (The uncommon
> admission is that we have no easy way to think about the later possibility).
All spikes might encode information. My thought is that the spatiotemporally
correlated spikes might be the bearers of spatial patterns of edges (by virtue
of their spatiotemporal volley patterns) and that the uncorrelated ones might
be bearers of properties such as local average luminance (through their firing
rates).
Some neurons clearly are coincidence detectors, and some are not. Even
neurons with many small inputs can fire with great precision if their
inputs are highly correlated. We see this in sustained chopper and onset
responders in the cochlear nucleus.
> Finally, the obsession with synchrony and cross correlations HAS produced
> two important advances:
>
> 1) it gave neurobiologists a safe stepping stone away from coding
> information in neuronal firing rates (safe in the sense that you can easily
> explain the idea in a paper in Nature or Science, and no real math is
> required).
>
> and 2) many neurobiologists are now turning to multi-single neuron
> recording techniques so they too can construct (and overinterprete) cross
> correlation histograms.
Absolutely! It's been a very, very positive development within the
neurosciences. I'm not knocking the focus on cross-correlations -- I
just don't see why they should be arbitrarily restricted to intrinsic
ones.
> >I think we desperately need to find a cure for myopia.
>
> sorry, that will have to wait until we understand consciousness. However,
> on that subject, I think we should focus on understanding things we can
> define first.
I'm sorry I was a bit flippant. Phase-locking is so important in many other
sensory systems. In audition, frequency discrimination follows the decline in
phase-locking as frequency increases. Pitches of complex tones follow the
interval patterns. The whole binaural time difference story and echolocation
all depend on phase-locked information. Electroception is the same way.
Flutter-vibration discrimination on the skin, same thing. Motion detection in
the fly. It just seems that there is a big blind-spot in vision research when
it comes to phase-locking. It seems almost like a taboo, a forbidden island
that the natives simply won't go to. Is it the third-rail of vision research?
Just unsexy, boring? Or is it simply a naive, dumb idea that anybody can see
can't possibly be the case, for reasons X, Y, and Z. If there are good
reasons against it, I want to know what they are, so that I can stop deluding
myself. If there really aren't good reasons, then someone should seriously
consider these possibilities.
This has been helpful -- thanks.
Peter Cariani