Probabilistic processing of sequential auditory information in noise
Abstract
Humans can typically follow and understand structured auditory information, such as speech or music, even if it is embedded in a noisy background. However, existing models aiming to replicate this skill are still far behind of what most people can do without much effort. The lack of success in matching human performance is due to the fact that the principles of sound processing used by the brain are still not well understood. We offer a novel cognitive computational approach for tackling this task. The basic idea is that during listening, the brain probabilistically deduces the causes of the sounds based on its internal model of the full acoustic environment and on statistical regularities that appear in the acoustic input. Our key hypothesis is that the brain’s internal model represents the full distribution of the putative sources of the environment. This approach differs from the traditional view, which instead of probabilistic representation of all the underlying causes,relies on the analysis and template matching of feature co-occurrences when detecting or identifying sound sources. If our assumption is correct, then even in a noisy environment, the model formed by the brain remains robust and can track the same sources continuously. For testing this assumption, we will develop a method for forming an appropriate model of both the noise and the signal. We will build our computational model and then validate it in three different ways. First, we will compare the model predictions to data obtained in human behavioral experiments. Next, we will conduct neurophysiological experiments in freely moving animals to investigate the biological plausibility of our approach. Finally, we will compare the performance of our computational model to the best available models for auditory information processing to assess its feasibility for various applications. Building a biologically plausible model for sound processing not only offers an opportunity to gain better understanding about how our brain works, but it can also promote applications in several areas of everyday life, such as improving hearing aids and hearing protecting devices, building better telecommunication devices, making communication more robust between humans and robotic systems in everyday (noisy) situations, and enhance voice-controlled devices. All of these applications require separating certain types of sounds, such as music or a spoken command from environmental noise. The role of signal and noise can change in different contexts to allow the perceiver, whether human or machine, to focus successfully on the sounds of interest. During the course of the research, we will regularly publish the most recent version of the computational model and the results of our studies regarding the biological plausibility of the model (i.e. behavioral validation with humans, and neurophysiological validation with higherorder animals). We plan to submit at least one paper per year to peer-reviewed journals. The source code of the model will be made available for the scientific community. Preliminary results will be presented at relevant conferences (e.g., COSYNE, ASA), at least once per year. Finally, we will provide annual reports on our progress.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 07, 2019
- Source ID
- N629091912029
Entities
People
- Istvn Ulbert
Organizations
- Office of Naval Research
- Pázmány Péter Catholic University
- United States Navy