Predictive intelligence in hearing

Abstract

In this project, the PI explores the benefits of a predictive architecture for processing complex audio signals and specifically detection of important events in realistic soundscapes. The effort is structured around two main tasks: Task 1 develops a hierarchical predictive model of auditory cortex whereby multiple cortical layers operate as predictive models that guide processing based on internal models. The proposed system explores new ideas that diverge from current state-of-the-art audio systems based on convolutional neural network systems, specifically: (i) lifelong learning, (ii) causal inference, (iii) robustness and invariance. This scheme will be tested in a framework of audio event processing for surveillance and target detection to flag salient events. The project will particularly focus on the predictive learning capability of the system as well as its robustness under various testing conditions. The model will be informed by an experimental effort outlined in Task 2, where the PI explores specific parameters of the model in human listeners such as resolution of internal models along multiple acoustic dimensions and sensitivity of subjects to different contexts. The experimental behavioral and neural responses will refine the computational architecture as well as explore potential and limitations of the predictive hypothesis, at the center of this effort.

Document Details

Document Type
DoD Grant Award
Publication Date
Nov 26, 2019
Source ID
N000141912689

Entities

People

  • Mounya Elhilali

Organizations

  • Johns Hopkins University
  • Office of Naval Research
  • United States Navy

Tags

Readers

  • Computational Modeling and Simulation
  • Neural Network Machine Learning.
  • Speech Processing/Speech Recognition.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks