Hyperdimensional Brain-Inspired Learning for Phoneme Recognition With Large-Scale Inferior Colliculus Neural...

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Decoding neural activity from the Inferior Colliculus (IC) is crucial for understanding the auditory system.
• Existing methods often require large datasets and extensive fine-tuning due to noisy neural signals.
• Deep Neural Networks (DNNs) are widely used but can be computationally intensive.

Purpose of the Study:

• To develop a novel and highly efficient framework for phoneme recognition by decoding Inferior Colliculus (IC) neural activities.
• To leverage Hyperdimensional Computing (HDC) as an alternative to DNNs for efficient neural activity analysis.
• To deploy the HDC-based algorithm on Field Programmable Gate Arrays (FPGAs) for hardware acceleration.

Main Methods:

• A spatial and temporal-aware HDC encoder was developed to capture global and local patterns in neural activity.
• The HDC algorithm was implemented on a Field Programmable Gate Array (FPGA) platform for optimized speed.
• Inferior Colliculus (IC) neural activities were recorded from gerbils during phoneme playback for evaluation.

Main Results:

• The proposed HDC method demonstrated superior classification quality compared to baseline machine learning algorithms.
• The HDC framework achieved significant runtime speedups: up to 74× on CPU, 67× on GPU, and 210× on FPGA compared to ResNet.
• Accuracy improvements of up to 15% for consonant and 10% for vowel classification were observed compared to ResNet.

Conclusions:

• Brain-inspired HDC enables efficient encoding of IC neural activity for phoneme classification.
• The framework offers orders of magnitude runtime speedup while enhancing accuracy.
• The HDC-based approach is scalable, viable for real-world deployment, and offers fast training with improved quality.