Phase and Feedback in the Nonlinear Brain. Malcolm Slaney (IBM and Stanford) Hiroko Shiraiwa-Terasawa (Stanford) Regaip Sen (Stanford)

Transcription

1 Phase and Feedback in the Nonlinear Brain Malcolm Slaney (IBM and Stanford) Hiroko Shiraiwa-Terasawa (Stanford) Regaip Sen (Stanford) Auditory processing pre-cosyne workshop March 23, 2004

2 Simplistic Models Dehydrated cats and the application of Fourier analysis to hearing problems become more and more a handicap for research in hearing. von Bekesy Overly simplistic models are bad Wrong path For example, critical band filters Pitch and phase

3 Tony s Goals Linearity Taylor says everything is linear To a point Temporal codes I don t know the answer. Attentional effects Big problem.. Not today Phase perception

4 A Critique of Pure Audition Marr Pure vision Churchland A critique of pure vision Slaney A critique of pure audition Where are efferents?

5 Motivation S. Shamma. Speech processing in the auditory system Good Shihab R. Carlyon, S. Shamma. An account of monaural phase perception Bad Shihab

6 Nonlinear Cochlear Mechanics BM Amplitude (db) Constant SPL 33nm 80dB 60dB 40dB 20dB Frequency (khz) Mössbauer Data SPL Q 3 CF 80dB 1 10k 60dB k 40dB k 20dB k dB 40dB 60dB 20 80dB Frequency (khz) From Basilar membrane measurements and the travelling wave, by B.M. Johnstone, R. Patuzzi, and G. K. Yates, Hearing Research, 22 (1986)

7 Cat Vowels (/da/)

8 Goal Test phase perception models with more sophisticated cochlear models Auditory models Gammatone Gammatone with hair cell model Lyon s passive Ear STRF

9 Nonlinearities Capture effect Dominant frequency is represented A few db matter Stochastic resonance Sub-threshold signals entrain firings Temporal coding Non-linear mapping

10 Phase Models Within channel Auditory nerve channel Phase matters Between channels Phase does matter! Single sinusoids 2 octave separation Time Time

11 Phase Perception Model Short-term Within channel Sensitive to small phase change STRF fails Today Long term Speech-rate perception Large phase changes (relative to CF) STRF works great

12 Models Gammatone Critical band filters Psychoacoustic data Linear bandpass filters Meddis Hair Cell Nonlinear reservoir model

13 Models Lyon s Passive Ear Lowpass filters Transmission line Cascade Nonlinear automatic gain control (AGC) Passive

14 Models STRF Spatial temporal response function Based on Shamma s cortical data Fit spectrogram data Two-dimensional sinusoids Image domain

15 Patterson Phase Results Stimulus Constant phase shift per critical band Shifted phase gradually Result Small local changes Large global change Couldn t prove channel effect!

16 Patterson Stimuli Harmonic signal 31 harmonics Phase Alignment Cosine phase Random phase Alternating phase (APH)

17 Our Model Sound A Cochleagram Correlogram Cochleagram Cochleagram Memory? Comparison Comparison Sound B Cochleagram Correlogram Cochleagram Cochleagram Memory? Integrated meansquared error

18 Auditory Dimensions One Dimensional Pressure Time Cochlear Processing Two Dimensional Cochlear Place Time Correlogram Processing Three Dimensional Cochlear Place Autocorrelation Lag Time

19 Correlogram Structure Pitch Period Cochlear Place Autocorrelators Cochlea

20 Duda Tones Simple tone Harmonic complex Tone build up (slow and fast) Tone removal (from top and bottom) Wiggle

21 Cocheagram Correlogram 2 GFB: F0=62.5 LH=3 H= correlogram GFB Results APH Alternating phases are shifted Cochleagram Auditory nerve Error grows Temporal pattern Error falls sum square error sum square error sum square error aph degree 2 x 108 cochleagram MH aph degree 1.5 x 10-3 cochleagram LPE aph degree sum square error sum square error sum square error aph degree 10 x 106 correlogram MH aph degree 5 x 10-3 correlogram LPE aph degree

22 APH Results sum square error Correlogram --- Gammatone FB LH = 3 LH = 7 LH = 15 Test Error wrt phase Lowest harmonic Gammatone Too much variance No one threshold Lyon s Ear sum square error sum square error aph degrees 4 x 108 GFB + Meddis Hair Cell LH = 3 LH = 7 LH = aph degrees Lyon Passive Ear LH = LH = 7 LH = aph degrees

23 10 2 scaled, f = 62.5, 125 reference Threshold Gammatone No one threshold APH threshold 10 1 experiment f = 62.5 simulation f = 62.5 experiment f = 125 simulation f = lowest harmonics Lyon s Passive Ear Consistent results Depends on loudness (scaling) APH threshold reference experiment f = 62.5 simulation f = 62.5 experiment f = 125 simulation f = 125 unscaled, f = 62.5, lowest harmonics

24 Craig and Jeffress Two components Reverse polarity Task Detect difference Identify phase

25 Craig & Jeffress Results 6 x sum square error btw/ original & inversed φ = 0 φ = 45 φ = 90 φ = sum square error Cb level

26 To Do Improve results Tune model bandwidths Better cochlear/hair cell integration Verify two distinct regimes Break in performance Not

27 Conclusions Phase perception model Two different models Short-term within channel Long-term STRF? Better auditory model Less about critical bands