Dynamic selective auditory attention detection using RNN and reinforcement learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• The cocktail party phenomenon highlights the brain's ability to selectively focus on auditory stimuli amidst noise.
• Selective auditory attention detection (SAAD) is crucial for advancing brain-computer interfaces and auditory processing systems.

Purpose of the Study:

• To develop a dynamic system for real-time selective auditory attention detection.
• To model SAAD as a sequential decision-making problem solvable with advanced machine learning techniques.

Main Methods:

• A novel dynamic SAAD system was proposed, processing temporal signal evolution.
• The system was modeled as a sequential decision-making problem.
• Recurrent Neural Networks (RNNs) and reinforcement learning (Q-learning, deep Q-learning) were employed as the core methodologies.

Main Results:

• The deep Q-learning approach, utilizing an RNN agent, demonstrated superior performance.
• Achieved the highest classification accuracy at 94.2%.
• Exhibited the minimal detection delay compared to other dynamic learning approaches.

Conclusions:

• The proposed dynamic SAAD system effectively processes sequential auditory inputs.
• This method offers significant advantages for real-time attention detection, especially when listener attention shifts.
• The system holds potential for applications requiring dynamic auditory focus tracking in complex acoustic environments.