Re: Pitch neurons (was "Autocorrelation-like ...") (Peter Cariani )

Subject: Re: Pitch neurons (was "Autocorrelation-like ...")
From:    Peter Cariani  <peter(at)EPL.MEEI.HARVARD.EDU>
Date:    Thu, 6 Mar 2003 15:54:28 -0500

On Thursday, March 6, 2003, at 08:27 AM, Martin Braun wrote: > Peter Cariani wrote: > >> Martin: "No problem. F0 periods that arise in the frequency laminae of >> the partials are forwarded to the frequency laminae of F0." > >> if this were the case then there should be true pitch detectors that >> respond both to a pure tone and a complex with a missing F0 at the >> same >> frequency. Schwarz & Tomlinson (1990) looked hard in the awake monkey >> cortex without success. > > F0 pitch and pure-tone pitch are detected by the same neurons in the > low-frequency laminae of the midbrain. After that they are not > represented > by the same neurons in the cortex. This is no surprise at all. An F0 > comes > together with spectral information (timbre), a pure-tone doesn't. We > have > completely different inhibitory response chains for the two stimuli. > Remember that pure-tones do not occur in nature, and that our hearing > is in > no way adapted to them. Yes I agree that pure tones are highly artificial, and if anything our hearing is adapted to complex, broadband sounds. It would solve the problem if IC units with 300 Hz BFs also responded to AM tones with Fm's of 300 Hz and carriers at 1 kHz. Unfortunately, I know of no evidence for such responses (if you have seen such things or know of a suitable reference, I would be more than happy to disabuse myself of my mistaken belief that none have been seen. It would relieve me of a good deal of irritating cognitive dissonance I experience whenever I think about the central coding of pitch). I am very confident, however, that one would see interspike interval correlates of the 300 Hz periodicity in both cases (even) at the level of the IC (they are visible in Langner's beautiful spike raster plots). The problem with the interval representation at that level is that units respond with much lower discharge rates that in earlier stations, so that intervals at 1/F0 are not in huge abundance. This problem might be overcome by looking at patterns of longer intervals (n/F0) or by a volley principle in which across-neuron intervals are used. Chow(1951) estimated that there are 400,000 neurons in the macaque midbrain, as opposed to 30,000 in its auditory nerve -- perhaps what is going on is that the timing information is being more sparsely distributed across many more neurons as you go up the pathway. Why this should be, or what this says about the basic signal processing principles that are operant in the system is anyone's guess. We tend to think reflexively in terms of signal reductions (sequential hierarchies of feature detectors) rather than signal expansions and statistical mass-action schemes. Some of this is a holdover from AI machine vision processing strategies from the 1960's. > >> Martin: "We only hear two pitches, if there are strong timbre labels >> attached to them. These are decoded in the cortex, which then feeds >> back to the pitch neurons in the midbrain." > >> Simultaneously play two notes a whole step (or more) apart on the >> piano. Do you hear one or two notes? > > No note at all. Only a chord. I've never been accused of being an overly analytical listener. Yes I can hear the chord, but I can also tell that there are two notes and I can match the pitches of the notes. For chords, one can often hear out the pitches of the constituent notes (see Parncutt's book, which has pitch matches for chords in it, or the paper on tonal fusion by De Witt and Crowder "Oomph in Stumpf".) I agree that for some simple intervals (4ths, 5ths, octaves) the tones are more likely to fuse than others, but if you pick a major 2nd or a minor 3rd, I'll wager that most people will tell you that more than one note is being played. >> On what basis can you say that the notes MUST be separated in the >> cortex? How "anatomically and physiologically realistic" is this >> assertion? > > Timbre detection is based on memory reference. This is not possible > below the cortex. I think it's difficult to rule out subcortical timbre discrimination at this point. Why do you think it's not possible? The cortex does seem to be involved in fine grained hearing (see Tramo et al and Phillip's paper on word deafness) although it is far from clear to me whether there are fine representations of pitch and timbre there at the cortical level proper or whether these would exist in cortico-thalamic loops, or whether the cortico-thalamic system sets up the lower level stations to make fine-grained comparisons. In any of these cases cortical damage would impede fine grained pitch and timbre perception. My impression is that people and animals with extensive bilateral cortical lesions can still make coarse pitch and timbral judgements, although one never knows whether some primary auditory regions are spared or have reorganized over time (many are completely deaf for some period immediately after the precipitating event -- which argues that you need some functioning cortex to hear sounds in the first place. Question: is there an auditory analogue of Blindsight in which people with cortical lesions can localize and discriminate sounds in the absence of consciously hearing them?). Are there any fMRI studies out there that look at both subcortical and cortical activation during echoic and working memory fine pitch-related auditory tasks? These might shed some light on the problem of where fine-grained representations are stored. > >> But it still seems to me that even at the level of the IC, >> interval representations for low pitches still have more plausibility >> than those based on modulation-tuned pitch detectors > > Interval representations are the matter that must be detected by > SOMETHING. > This something are tuned pitch neurons. This is exactly what I meant about our fixation on single neurons and tuned feature detectors. There are other possible ways that neural networks can handle information through distributed representations and processings that do not require highly tuned units. The connectionists have various demonstration models of how this could conceivably be realized with Hopfield nets (rate-based codes), although the story gets very complicated if units are very coarsely tuned and have level-dependent and nonmonotonic responses. It's not clear to me whether anyone has demonstrated how these mechanisms would work in practice, using the messy kinds of responses that the physiologists actually see. I have found that feedforward timing nets in principle can be used to beat interval distributions against each other in a coincidence-delay network to compare interval distributions in a manner that could operate on sparse-distributed temporal patterns, but this is also still only a demonstration of functional principle rather than a concrete model of the behavior of a particular set of neurons. So, there are many possible ways to skin a cat as they say, and the limits largely reside in our own imaginations. This is why it is so dangerous and destructive to rule out a whole class of possible neural explanations (interval representations) because we don't yet know how they are processed. Even though I am highly, highly critical of central rate codes and am critical of these modulation-based representations, we still cannot rule them out of hand -- maybe there are ways that we/I haven't thought of in which they operate. Similar kinds of dismissive thinking set back the theory of continental drift by half a century -- the reigning geologists at the time could not imagine a motor mechanism which would propel the continents, and the hypothesis was ridiculed for decades until World War II and the discovery of sea floor spreading forced that community to re-evaluate it. My father was a geologist and as a young boy I remember the controversy, even after a great deal of supporting evidence had been amassed. Continents that move is a hard thing to get used to if you've always thought of them in other terms. Preposterous! >> Martin: "Level stability is provided by lamina-based lateral >> inhibition in the midbrain." > >> Show me the neural data. I want to see MTFs and pitch representations >> that are every bit as good at 90 dB SPL as they are at 40 or 50 dB. > > MTFs [modulation transfer functions] do not occur with a piano tone. > With a piano tone you have beats between the partials in the > middle-frequency laminae of the midbrain, which are then detected by > tuned pitch neurons. Sound level does not affect either the partial > frequencies or the beat frequencies. It only affects the mutual > inhibition of partial frequencies, and thus strength of pitch, but not > pitch height. Schouten and de Boer's experiments on inharmonic complex tones pretty much destroyed theories of F0-pitch based on neighboring harmonics that beat. This is one of the strong demonstrations that argues that low (F0) pitch (of resolved harmonics) does not depend on envelope periodicity per se, and that the fine structure of the waveform must be analyzed in a manner that is similar to autocorrelation. Spectrally, it means that you can't simply infer F0 from harmonic spacings -- a harmonic pattern analysis must be carried out. I know of no good physiological evidence for any central harmonic spectral pattern analysis (excluding bats and combination-sensitive units with BFs > 5 kHz). >> Martin: "Period detectors register these pitch shifts, as calculated >> many years ago." > >> Only if they are based on autocorrelation or something roughly >> equivalent. Modulation detectors won't work ... > > See last answer. Also see what I wrote on beats in the midbrain > (Braun, 1999). You will then see that pitch neurons that function as > tuned resonators can do all what autocorrelators can. Subcellular time > constants are solutions of biology, whereas autocorrelation circuits > are solutions of man-made electronics. I will try to look again at your 1999 paper. Maybe our quibble is semantic. If you see pitch neurons that "do all what autocorrelators can", then functionally they are autocorrelators, and they will have properties that are associated with those (e.g. comb filter characteristics). Autocorrelations differ from renewal processes in that they don't reset, and this in turn means that they handle competing and interleaved temporal patterns without much destructive interference. Autocorrelators are not tied to any particular man-made electronic circuits -- they do require delay operations but there are many ways besides delay lines to realize those, although tapped delay lines and axons with collaterals are obvious analogues. A decade ago, I initially thought that "intrinsic oscillations" in CN choppers might be able to implement a Licklider architecture using cellular autocorrelators, but when I looked at chopper responses to periodic complex stimuli, yes there was some very coarse F0 tuning, but the interval information was still much, much better and as in the auditory nerve, it follows the psychophysics. More power to you if you can figure out a way to make a cellular autocorrelator it that is compatible with the observed neural responses and the neuroanatomy. I would not even rule out the possibility of sub-cellular tuned resonances in particular axons and dendrites (see the Waxman book). It's possible that we may be able to get something like an autocorrelation out of the bandpass MTF neural elements in the pathway or from combinations of elements, and I think you should keep trying -- it would be a huge step forward if you could do it. Just don't be so blindly dismissive of other possibilities in the meantime. --Peter Cariani Cariani P (2001) Neural timing nets. Neural Networks 14:737-753. Chow KL (1951) Numerical estimates of the auditory central nervous system of the rhesus monkey. J Comp Neurol 95:159-175. DeWitt LA, Crowder RG (1987) Tonal fusion of consonant musical intervals: The oomph in Stumpf. Perception and Psychophysics 41:73-84. Parncutt R (1989) Harmony: A Psychoacoustical Approach. Berlin: Springer-Verlag. Phillips DP, Farmer ME (1990) Acquired word deafness, and the temporal grain of sound representation in the primary auditory cortex. Behavioural Brain Research 40:85-94. Tramo MJ, Shah GD, Braida LD (2002) The functional role of auditory cortex in frequency processing and pitch perception. J Neurophysiol 87:122-139. Waxman SG (1978) Regional differentiation of the axon: A review with special reference to the concept of the multiplex neuron. In: Physiology and Pathobiology of Axons (Waxman SG, ed). New York: Raven Press.

This message came from the mail archive
maintained by:
DAn Ellis <>
Electrical Engineering Dept., Columbia University