Subject: Re: pitch neurons (2) From: Peter Cariani <peter(at)EPL.MEEI.HARVARD.EDU> Date: Wed, 9 Oct 2002 18:24:59 -0400--Apple-Mail-2-60524677 Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset=US-ASCII; format=flowed Comment # 2 (Wednesday, Oct. 9) Israel Nelken wrote: Martin and list, I think that the idea that neurons in the inferior colliculus are sensitive to different periodic modulations of their input is generally accepted. Many neuronal mechanisms would do that - not only intrinsic oscillations. I don't think this is really the issue about which I feel uncomfortable. The point is that these neurons mostly respond to sounds with energy content within their tuning curve (with a lot of fine print attached, I agree), and therefore cannot be interpreted as 'pitch neurons'. When you say 'pitch neuron', I understand that you are looking for a neuron that should respond to pitch irrespective of the underlying physical structure of the stimulus - missing fundamental or not, iterated ripple noise, high-pass filtered click trains, binaural or monaural, and so on and so forth. I don't think that there's any data in the literature (including the very nice Fishman et al. paper that you referred to in a previous message) that even come close. It is precisely this high abstraction level of the pitch, its independence of so many of the physical attributes of the underlying sounds, that to my opinion argues for a high level of processing for the common denominator of all of these sounds. This however does not preclude a partial extraction of some of the features of such sounds earlier (e.g. periodicity at the level of the IC). =================================================================== Hi Eli, Dennis, Martin, others, The big point is that the all-order interval representations at the level of auditory nerve, brainstem, midbrain DO account for these pitch equivalences. There is no need to pass the buck up to the cortex, and to invoke complicated, unspecified mechanisms to generate all these pitch equivalences if there is some mechanism by which the interval information is analyzed. I don't know what that mechanism is (I have working hypotheses), but I believe that it is absolutely fundamental to understanding auditory form perception (speech, music, sound recognition). It is only with an interval code for a front-end that one gets a representation that is highly invariant with respect to sound pressure level, i.e. in a manner that behaves like most auditory form percepts. Yes, it is possible, always, to postulate separate subpopulations to handle each sort of stimulus/stimulus parameter, but stringing all the different mechanisms together in an ad hoc manner and without any kind of hypothesis for the integrative mechanisms (why there should be pitch equivalence in the first place, how this disparate information is put together) seems like an evasion to me. It dismisses the problem rather than confronting it. The reason that these sounds look so diverse and disparate has to do with viewing them from the particular perspective of Fourier analysis -- they all have deep underlying common properties if you look at them through the lens of autocorrelation (or other alternative representational schemes). We all agree that there are neurons in the IC that have bandpass MTF functions. This is possibly the best theory that we have at the moment for some complex tone pitches, but it is a very weak explanation, with many problems: 1) IC MTF functions generallyhave broad tunings to begin with when compared with pitch discriminations, 2) they flatten out at higher SPLs when pitch percepts remain highly salient, 3) Bandpass MTFs cannot explain pitch equivalence between pure and complex tones (why would it be that all auditory systems down to goldfish do this?) 4) they cannot explain the pitch shifts of inharmonic AM tones (first effect of pitch shift, resolved harmonics, de Boer's rule), nor can they explain how it is that we can hear the pitches associated with both of 2 concurrent notes on the piano that are a few semitones apart (these interact destructively in an MTF-based representation) At the cortex, the coding of pitch is still very murky -- it is possible that no F0-sensitive neurons were found in Schwarz & Tomlinson's study because the monkeys weren't paying attention, and attention is a potentially important, though usually not decisive, factor. However, the missing-F0 complexes that they used evoke pitches (in us) that are comparable in salience to pure tones. I think it's highly unlikely the monkeys were only paying attention to the pure tones, for which S & T did obtain tuned responses (and one finds these in anesthetized preps as well). Part of the problem is that people tend to think that pure tones are simpler stimuli than complex tones (true in a Fourier description but not nec. in other representations). (As Dennis well appreciates, the problem of representing pure tones in the cortex gets very complicated once one has to account for the stability of the representation over different SPLs and source locations.) (I wonder if anyone has looked at reaction times for same-different pitches for pure tones & complex tones. This might bear on some of the questions related to whether processing for pure tones is less complex than for missing-F0 stimuli). We need to expand the range of possible kinds of mechanisms that are considered. We implicitly first want to see nice, spatially ordered maps of periodicity-tuned elements. Failing that, we would like to see mosaics of units with the right kinds of response properties that would account for pitch (comb filter tunings, temporal autocorrelators). Failing that, we might think that there is some kind of covert, distributed processing going on in which combinations of units are analyzed (how this leads to the low-dimensional structure of pitch space and pitch classes is quite unclear). Failing that, maybe we should consider that the representation lies in relations between spikes rather than in patterns of units being activated -- i.e. the whole idea of feature detectors is suspect. At some point when all models fail, we must go back, explicitly identify our basic assumptions, question them, and come up with alternatives. -- Peter Cariani --Apple-Mail-2-60524677 Content-Transfer-Encoding: 7bit Content-Type: text/enriched; charset=US-ASCII Comment # 2 (Wednesday, Oct. 9) Israel Nelken wrote: <color><param>0000,0000,DEDE</param>Martin and list, I think that the idea that neurons in the inferior colliculus are sensitive to different periodic modulations of their input is generally accepted. Many neuronal mechanisms would do that - not only intrinsic oscillations. I don't think this is really the issue about which I feel uncomfortable. The point is that these neurons mostly respond to sounds with energy content within their tuning curve (with a lot of fine print attached, I agree), and therefore cannot be interpreted as 'pitch neurons'. When you say 'pitch neuron', I understand that you are looking for a neuron that should respond to pitch irrespective of the underlying physical structure of the stimulus - missing fundamental or not, iterated ripple noise, high-pass filtered click trains, binaural or monaural, and so on and so forth. I don't think that there's any data in the literature (including the very nice Fishman et al. paper that you referred to in a previous message) that even come close. It is precisely this high abstraction level of the pitch, its independence of so many of the physical attributes of the underlying sounds, that to my opinion argues for a high level of processing for the common denominator of all of these sounds. This however does not preclude a partial extraction of some of the features of such sounds earlier (e.g. periodicity at the level of the IC). </color> =================================================================== Hi Eli, Dennis, Martin, others, The big point is that the all-order interval representations at the level of auditory nerve, brainstem, midbrain DO account for these pitch equivalences. There is no need to pass the buck up to the cortex, and to invoke complicated, unspecified mechanisms to generate all these pitch equivalences if there is some mechanism by which the interval information is analyzed. I don't know what that mechanism is (I have working hypotheses), but I believe that it is absolutely fundamental to understanding auditory form perception (speech, music, sound recognition). It is only with an interval code for a front-end that one gets a representation that is highly invariant with respect to sound pressure level, i.e. in a manner that behaves like most auditory form percepts. Yes, it is possible, always, to postulate separate subpopulations to handle each sort of stimulus/stimulus parameter, but stringing all the different mechanisms together in an ad hoc manner and without any kind of hypothesis for the integrative mechanisms (why there should be pitch equivalence in the first place, how this disparate information is put together) seems like an evasion to me. It dismisses the problem rather than confronting it. The reason that these sounds look so diverse and disparate has to do with viewing them from the particular perspective of Fourier analysis -- they all have deep underlying common properties if you look at them through the lens of autocorrelation (or other alternative representational schemes). We all agree that there are neurons in the IC that have bandpass MTF functions. This is possibly the best theory that we have at the moment for some complex tone pitches, but it is a very weak explanation, with many problems: 1) IC MTF functions generallyhave broad tunings to begin with when compared with pitch discriminations, 2) they flatten out at higher SPLs when pitch percepts remain highly salient, 3) Bandpass MTFs cannot explain pitch equivalence between pure and complex tones (why would it be that all auditory systems down to goldfish do this?) 4) they cannot explain the pitch shifts of inharmonic AM tones (first effect of pitch shift, resolved harmonics, de Boer's rule), nor can they explain how it is that we can hear the pitches associated with both of 2 concurrent notes on the piano that are a few semitones apart (these interact destructively in an MTF-based representation) At the cortex, the coding of pitch is still very murky -- it is possible that no F0-sensitive neurons were found in Schwarz & Tomlinson's study because the monkeys weren't paying attention, and attention is a potentially important, though usually not decisive, factor. However, the missing-F0 complexes that they used evoke pitches (in us) that are comparable in salience to pure tones. I think it's highly unlikely the monkeys were only paying attention to the pure tones, for which S & T did obtain tuned responses (and one finds these in anesthetized preps as well). Part of the problem is that people tend to think that pure tones are simpler stimuli than complex tones (true in a Fourier description but not nec. in other representations). (As Dennis well appreciates, the problem of representing pure tones in the cortex gets very complicated once one has to account for the stability of the representation over different SPLs and source locations.) (I wonder if anyone has looked at reaction times for same-different pitches for pure tones & complex tones. This might bear on some of the questions related to whether processing for pure tones is less complex than for missing-F0 stimuli). We need to expand the range of possible kinds of mechanisms that are considered. We implicitly first want to see nice, spatially ordered maps of periodicity-tuned elements. Failing that, we would like to see mosaics of units with the right kinds of response properties that would account for pitch (comb filter tunings, temporal autocorrelators). Failing that, we might think that there is some kind of covert, distributed processing going on in which combinations of units are analyzed (how this leads to the low-dimensional structure of pitch space and pitch classes is quite unclear). Failing that, maybe we should consider that the representation lies in relations between spikes rather than in patterns of units being activated -- i.e. the whole idea of feature detectors is suspect. At some point when all models fail, we must go back, explicitly identify our basic assumptions, question them, and come up with alternatives. -- Peter Cariani --Apple-Mail-2-60524677--