1. An auditory system modeling in sound localization
EE381K MDDSP
May
Yul Young Park

2. Previous Talk Sound Localization motivations:
virtual 3D sound, game, sonar, microphone array, objective sound quality evaluation, education, etc
Goal: estimate correct azimuth and elevation (distance -> ignored)
Conventional approach: neural network [Neti&Young,1992] vs probabilistic estimator [Lim&Duda,1994, Chau&Dudda,1995, Martn,1995]
Suggested method: improve neural network model by evolutionary computation [Stanley,2002]

3. example block diagram [Lim&Duda,1994]

5. Neural Network training data, test data
- network output:
- desired output:
- mean square error of network:
gradient descent method:
decrease the error by adjusting
training data, test data
issues: local minimum, generalization, data representation

6. Evolutionary Computation [stanley,2002]
Local Minimum, partly Generalization
Evolution on Node and connection genes: population -> evaluation (fitness) -> reproduce (ranking/sampling/crossover or mutation/speciation)
Innovation record: historical marking -> easy topology analysis

7. Cochlear & Neural Signal Model
Where I am
HRTF
Cochlear & Neural Signal Model
Pre-Processing
Neuroevolution
Elevation: -40°~90°, 10° increment
Azimuth: 5°~30° at each elevation
Total 710 points max.
gammatone filterbank &half wave rectification
64 ch. along 500Hz ~22.05kHz
ITD: cross-correlation
IID: subtraction
population size
fitness evaluation
computation issues: data representation (int/real, bipolar/unipolar), dataset size (resolution/time)