Listen to the Brain-Auditory Sound Source Localization in Neuromorphic Computing Architectures

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Auditory Processing

Background:

• Conventional sensory processing is inefficient due to uniform sampling, leading to redundant data.
• Neuromorphic computing offers an alternative by mimicking biological systems with event-based hardware.
• Sound source localization (SSL) is a key auditory processing task where efficiency gains are desirable.

Purpose of the Study:

• To propose a generic approach for mapping biologically inspired neural networks to neuromorphic hardware for SSL.
• To demonstrate the implementation of a neural SSL model on two distinct neuromorphic platforms.
• To evaluate the performance of the implemented models using synthetic and real-world auditory data.

Main Methods:

• Modeled neural mechanisms of SSL based on interaural level difference (ILD).
• Identified and transformed computational motifs into spike-based components for neuromorphic implementation.
• Mapped the SSL model to IBM TrueNorth and SpiNNaker neuromorphic hardware platforms.

Main Results:

• Both neuromorphic implementations achieved 100% accuracy for synthetic sound stimuli.
• Real-world experiments showed varying performance: TrueNorth (78% accuracy, 41° RMSE, 18° MAE) and SpiNNaker (13% accuracy, 39° RMSE, 29° MAE).
• The study successfully demonstrated the same SSL model's implementation across different neuromorphic architectures.

Conclusions:

• The proposed mapping approach facilitates hardware-independent neural SSL.
• This work paves the way for efficient, biologically inspired auditory processing on diverse neuromorphic systems.
• Further optimization is needed for real-world performance, particularly on the SpiNNaker platform.