Exploring neural oscillations during speech perception via surrogate gradient spiking neural networks

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neuromorphic engineering

Background:

• Cognitive processes require large-scale neural dynamic models.
• Existing models often lack physiological inspiration and scalability.
• Understanding neural synchronization in auditory pathways is key.

Purpose of the Study:

• To develop a scalable, physiologically inspired speech recognition architecture.
• To investigate the emergence of neural oscillations via gradient descent training.
• To analyze the role of feedback mechanisms in neural synchronization and performance.

Main Methods:

• Physiologically inspired spiking neural network architecture.
• End-to-end gradient descent training.
• Analysis of cross-frequency couplings and neural activity patterns.
• Evaluation of feedback mechanisms (spike frequency adaptation, recurrent connections).

Main Results:

• Emergence of neural oscillations during speech processing in the spiking neural network.
• Significant cross-frequency couplings observed within and across network layers.
• Absence of these interactions during background noise processing.
• Demonstrated inhibitory role of feedback mechanisms in synchronizing neural activity.

Conclusions:

• The architecture successfully replicates neural dynamics and exhibits emergent oscillations.
• Feedback mechanisms are vital for regulating neural synchrony and enhancing speech recognition.
• The model offers insights into auditory pathway synchronization and efficient neuromorphic computing.