Learnable axonal delay in spiking neural networks improves spoken word recognition

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Speech Processing

Background:

• Spiking neural networks (SNNs) offer biologically plausible models for auditory processing, crucial for tasks requiring precise temporal information.
• The temporal complexity of spike sequences has historically limited SNN performance compared to artificial neural networks (ANNs).

Purpose of the Study:

• To enhance the performance of SNNs in spoken word recognition by addressing challenges in spike-timing configuration and network architecture.
• To develop a novel SNN architecture that achieves state-of-the-art results on challenging speech benchmarks.

Main Methods:

• Implemented a learnable axonal delay module within the SNN architecture.
• Integrated local skip-connections to facilitate information flow and gradient propagation.
• Introduced an auxiliary loss term to improve model accuracy and stability.

Main Results:

• Achieved state-of-the-art performance on spoken word recognition benchmarks (NTIDIDIGITS and SHD).
• Demonstrated significant performance improvements: 14% on NTIDIDIGITS and 18% on SHD with the delay module.
• Outperformed recurrent and convolutional neural networks using substantially fewer parameters (10x for NTIDIDIGITS, 7x for SHD).

Conclusions:

• Learnable axonal delays and local skip-connections are effective in advancing SNN performance for speech recognition.
• The proposed SNN approach offers a computationally efficient and biologically inspired alternative to conventional deep learning models for auditory tasks.