Leslie Famularo on Differentiating and Optimizing an Auditory Model
Date:
Fri, 08/09/2024 - 12:00am - 12:00pm
Location:
CCRMA Seminar Room
Event Type:
Hearing Seminar 
One of the shortcomings of current AI work is the inability to tie the results back to known physics. This is useful both to help explain the results, but also to constrain the optimal solution to known physical properties of the system. Neural networks are hard. They are big, often times the result is inscruttable. What can be done?
New software paradigms such as JAX and PyTorch allow one to specify arbitrary computations in a way that can be differentiated. And if we can differentiate a function we can optimize it. Hurray. How can we express an auditory model in a differentiable fashion?
Leslie Famularo is here in Silicon Valley this week and I’ve asked her to come to the Hearing Seminar to discuss her approach to optimizing auditory models. I’m sure we’ll have a bit of a discussion on the importance of performance, damn the physics, versus physically realizable models.
Who: Leslie Famularo (UMd)
When: 10:30AM Friday August 9th, 2024
What: A differentiable model of hearing for audio processing
Where: CCRMA Seminar Room, Top Floor of the Knoll at Stanford
Why: The cool kids know how to differentiate. :-)
If you missed the excellent discussion we had about the history of speech processing for Cochlear Implants, then please check out the video. Pretty amazing how hard it was back in the day, and how well it works now.
https://youtu.be/9hoY24bVTZw
See you at CCRMA on Friday morning
New software paradigms such as JAX and PyTorch allow one to specify arbitrary computations in a way that can be differentiated. And if we can differentiate a function we can optimize it. Hurray. How can we express an auditory model in a differentiable fashion?
Leslie Famularo is here in Silicon Valley this week and I’ve asked her to come to the Hearing Seminar to discuss her approach to optimizing auditory models. I’m sure we’ll have a bit of a discussion on the importance of performance, damn the physics, versus physically realizable models.
Who: Leslie Famularo (UMd)
When: 10:30AM Friday August 9th, 2024
What: A differentiable model of hearing for audio processing
Where: CCRMA Seminar Room, Top Floor of the Knoll at Stanford
Why: The cool kids know how to differentiate. :-)
If you missed the excellent discussion we had about the history of speech processing for Cochlear Implants, then please check out the video. Pretty amazing how hard it was back in the day, and how well it works now.
https://youtu.be/9hoY24bVTZw
See you at CCRMA on Friday morning
A differentiable model of hearing for audio processing
With the increase in end-to-end methods in audio processing, models of human auditory perception have been less favored in state-of-the-art methods. On the other hand, end-to-end methods suffer from its demand for data and computing resources, and are not explainable. My PhD research proposes that introducing auditory neuroscience models into machine learning for audio through differentiable programming can increase robustness and explainability. In this talk, I will introduce our ongoing effort to create a differentiable model of peripheral hearing and cortical processing, along with specific applications to phoneme recognition and source separation. Additionally, I will present some proposed work related to personalized headphone/hearing aid fitting.
Leslie Famularo is a 5th year PhD student in Neuroscience and Cognitive Science at the University of Maryland, advised by Drs. Ramani Duraiswami and Naomi Feldman. She also finished a M.S. in Computer Science while pursuing her PhD. Her research interests center on the intersections between perception/neuroscience and machine learning, both improving AI/ML using neuroscience knowledge and using AI/ML tools to create models of perception and learning.
FREE
Open to the Public
