Explore the intricacies of auditory cortical computation in this comprehensive lecture by Josh McDermott from MIT. Delve into the complexities of auditory input processing, starting with an examination of typical sound stimuli and the listener's interest in identifying their worldly origins. Analyze the standard peripheral auditory model and the linear spectrotemporal filtering approach in auditory cortex modeling. Investigate the potential for improved models through task-based system training, while acknowledging inherent limitations. Compare behavioral aspects such as speech recognition in noisy environments, sound localization, and pitch perception between computational neural networks and human performance. Discover how task performance correlates with the ability to predict neural responses and examine example metamers from various convolutional stages. Gain valuable insights into the ongoing debate between timing and "place" information in auditory processing.
Overview
Syllabus
Intro
Consider an example of typical auditory input
The listener is interested in what happened in the world to cause the sound
Standard peripheral auditory model
Standard model of auditory cortex linear spectrotemporal filtering
Can we obtain better models by training systems to perform tasks?
Some obvious limitations
Behavioral comparison: Speech recognition in background noise
Behavioral comparison: CNN & humans on same task
Behavioral comparison: Sound localization
Network learns ear-specific cues to elevation, like humans
Behavioral comparison: Pitch perception
Longstanding controversy over timing vs. "place" information
Task performance correlates strongly with ability to predict neural responses
Example metamers from each convolutional stage
Summary
Taught by
MITCBMM
