Search Papers | Poster Sessions | All Posters
Poster A80 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Psychoacoustic phenomena explained by auditory task optimization
Mark Saddler1,2 (), Torsten Dau1, Josh McDermott2; 1DTU, 2MIT
Artificial neural networks optimized for ecological tasks have emerged as leading models of sensory systems. Models optimized separately for sound localization and recognition tasks account for a range of human auditory behaviors, but it has remained unclear whether a single model could account for behaviors in both types of tasks. We optimized a model to jointly localize and recognize sounds from simulated auditory nerve input. The resulting multi-task model reproduced a range of human speech recognition effects related to noise, reverberation, and spatial separation. We also trained linear classifiers to perform simple psychoacoustic tasks using the model’s internal representations. The learned model features produced human-like patterns of psychoacoustic judgments. The results provide further evidence that many aspects of human hearing can be understood as optimized solutions to ecological tasks.
Keywords: auditory perception psychophysics deep neural network multi-task learning