Memristor-based analogue computing for brain-inspired sound localization with in situ training

Area of Science:

• Neuroscience
• Computer Engineering
• Materials Science

Background:

• The human nervous system processes sensory information analogically and efficiently.
• Sound localization is a key brain function, vital for virtual auditory systems.
• Memristor arrays face challenges in hardware demonstrations for localization tasks due to output neuron contributions.

Purpose of the Study:

• To experimentally demonstrate in-situ learning for sound localization using a memristor array.
• To introduce and validate a multi-threshold-update scheme for improved performance.
• To advance the development of low-energy, high-performance memristor-based auditory localization systems.

Main Methods:

• Utilized a 1K analogue memristor array for experimental demonstration.
• Implemented a proposed multi-threshold-update scheme for in-situ learning.
• Evaluated training accuracy and energy consumption against existing methods.

Main Results:

• Successfully demonstrated in-situ learning of sound localization.
• Achieved a ~45.7% improvement in training accuracy compared to existing methods.
• Reduced energy consumption by approximately 184 times relative to previous work.

Conclusions:

• The multi-threshold-update scheme enables efficient in-situ learning for sound localization in memristor arrays.
• This approach significantly enhances training accuracy and reduces energy consumption.
• Represents a substantial step towards practical memristor-based auditory localization systems.