Transcript es2005 5402

Studies of Information Coding
in the Auditory Nerve
Physiology
Modeling
Psychophysics
Laurel H. Carney
Syracuse University
Institute for Sensory Research
Departments of Biomedical & Chemical Engineering and
Electrical Engineering & Computer Science
Outline
• Background - Siebert’s Analytical Studies of
Coding in the Auditory-Nerve
– Rate-Place(frequency) Model
– All-Information Model (Temporal & Rate cues)
• Extending the approach with Computation
• Examples:
– Freq. Discrimination (tones)
– “Formant” Freq. Discrimination
– Level Discrimination (tones)
• From Ideal Observers to more ‘Realistic’ Models:
– Coincidence Detectors for Level Decoding that
combine Rate & Temporal Information
Coding of Sound in Auditory Nerve:
Tuning Curves suggest a “Rate vs. Place” code
But…..
(After Kiang)
Saturation of rate is a problem for
the rate-place encoding scheme
Note:
as Rate ’s,
Variability ’s
Rate is not adequate to encode stimulus energy at the fiber’s CF.
Additional
information
is present
in the
timing of
AN
responses
Siebert (‘68,‘70): Can the Limits of Human
Perception for Frequency and Level be explained
by basic properties of Auditory-Nerve responses?
- Analytical model
- Simple tuning; Place map
Log Frequency
- Saturating rate-level functions
- Steady-state responses
- Phaselocking included (results
limited to low freqs)
- Random nature of AN
responses described by Nonhomogeneous
Poisson
process
Siebert’s Approach
Applied to
Frequency Discrimination
(from Heinz et al., 2001,
Neural Computation)
Use of Cramer-Rao Bound to estimate jnd
Lower Bound on variance of frequency
estimate [based on r(t)] depends on rate
(Poisson assumption) and on partial
derivative of rate w.r.t. parameter of
interest
[1/ variance] can be summed over population of fibers
(assuming independence between fibers)
Discrimination Threshold, or Just-Noticeable
Difference (jnd), corresponds to difference in
parameter of interest that equals standard deviation.
Comparison of Siebert’s Predictions to Human Performance:
Frequency Discrimination
Rate-Place
All-Info
Rate-Place
All-Info
(after Heinz et al., 2001,
Neural Computation)
Siebert’s (‘68,’70) results suggest Rate-Place model for
Human Frequency Discrimination at low frequencies.
But Frequency discrimination
gets Worse at High Freqs,
and Rate-Place model doesn’t !
- Siebert’s analysis was limited by
simple peripheral model.
- Can extend the approach using
a Computational Model for AN
fibers (Heinz et al., 2001) :
-Allows phase-locking to rolloff
accurately vs. Freq.
Does a more complete AN model
change our conclusion?
Rate-Place
All-Info
Detailed AN
response
properties
included in
Computational
AN model:
- Phase-locking
- Onset/offsets
Comparison of Siebert’s Predictions to Human Performance:
Frequency Discrimination
Rate-Place
All-Info
Rate-Place
All-Info
(after Heinz et al., 2001,
Neural Computation)
Summary of Heinz et al.’s results:
• All-Info model matches trends in Human data,
for Frequency (and Level) Discrimination.
• Rate-Place model can’t explain Freq Discrim at high freqs.
• But, Thresholds of Optimal model are too low.
Optimal models help identify cues that are consistent with
overall performance of listeners.
More realistic (sub-optimal) processing mechanisms will have
elevated thresholds that do a better job of predicting both the
trends and absolute thresholds of human performance.
Extension of Siebert/Heinz
approach to Complex Stimuli
• Modeling Discrimination
of Center Frequency of
Formant-like Harmonic
Complexes (Tan & Carney, JASA, 2005)
Results for Human
Listeners (Lyzenga & Horst, 1995)
Center Freq Discrim JNDs
for 3 spectral slopes
Lowest thresholds
are for Center Freqs
between Harmonics
Highest thresholds
are for Center Freqs
at Harmonic freqs
Energy-based
model predicts
the opposite
Center Frequency (Hz)
AN Models require Timing Info to
Predict correct Threshold Trends
AN Model based on Timing Info in Small
# of Fibers Provides Best Predictions
AN Population
Model Predictions
• For Harmonic Complexes Timing Information is
required to predict trends in human performance
• But, Optimal Detector uses all timing information What aspect of ‘timing’ is critical for these results?
• Can use Sub-Optimal Detectors to explore
different aspects of timing:
e.g. Across-fiber timing (spatio-temporal patterns)
vs. Within-fiber timing patterns (intervals)
Level Coding in the Auditory Nerve based on
Sub-Optimal Processing: Coincidence-Detection
• Level-dependent tuning of Basilar Membrane
results in level-dependent timing of AN responses
(Anderson et al., 1971).
• At low frequencies, this neural cue may contribute
to level coding over a wide dynamic range.
• At high frequencies, level-dependent gain results in
wide dynamic ranges of AN fibers.
• Cross-frequency Coincidence Detection can take
advantage of both rate and timing cues.
Timing (phase)
of AN spikes
varies
systematically
with Level
(Response Area
from Anderson et al., 1971,
J. Acoust.Soc.Am.)
Level-dependent BW, Gain, & Phase are
included in computational AN model
Low SPL
Magnitude
Hi SPL
Phase
(Zhang et al., 2001, J. Acoust. Soc. Am.)
Nonlinear Auditory-Nerve model has:
- Nonlinear timing (dominant @ low Frequencies)
- Wide-dynamic ranges (dominant @ high Frequencies)
(Heinz et al., ARLO, 2001)
Low SPL
Magnitude
Phase
Hi SPL
Coincidence Detector
CDs are sensitive to rate and/or timing!
Level Discrimination Predictions based on
Coincidence Detection (CD) Model
1 kHz
10 kHz
Inputs to CD from Nonlinear Computational AN model
Decision variable based on Rate of CD
---Nonlinear Temporal cues important at low frequencies
---Wide-dynamic-range rate-level functions important at high
(Heinz et al., 2001, J. Acoust. Soc. Am.)
frequencies
Conclusions:
• Can quantify info in computational Auditory-Nerve model
response and compare to psychophysical performance.
• Combined Rate and Temporal info (“All-Info”) explains
trends in listeners across a wide range of tone frequencies
and levels, and for harmonic complex freq discrim task.
• Coincidence Detection (CD) is a simple mechanism for
decoding Temporal and/or Rate info.
• CD is consistent with trends & absolute thresholds of
Human Performance for Level Discrimination.
• CD does not explain performance in Harmonic
Complex task. Prelim results suggest that an intervalbased strategy to coding Instantaneous Frequency
or a modulation-based strategy are more promising.
Collaborators:
Michael Heinz - PhD 2000, HST-MIT; now at Hopkins
Steve Colburn - Dept. of Biomedical Engr., Boston University
Qing Tan - PhD, 2003 Boston University
• Supported by NIH-NIDCD, NSF, & The Gerber Fund
•NOTE: Code and papers are available at:
http://web.syr.edu/~lacarney/