Spiking mode-based neural networks.

Area of Science:

• Computational neuroscience
• Neuromorphic engineering

Background:

• Spiking neural networks (SNNs) are crucial for brain-inspired computing and understanding neural circuits.
• Training large-scale SNNs is computationally expensive, and the learned weights obscure circuit mechanisms.

Purpose of the Study:

• To propose a novel training protocol for SNNs that addresses high training costs and lack of transparency.
• To enable a more interpretable understanding of neural circuit dynamics.

Main Methods:

• Introduced a spiking mode-based training protocol using a Hopfield-like decomposition of the recurrent weight matrix.
• Represented the weight matrix as a multiplication of input modes, output modes, and a score matrix.
• Trained networks in the reduced mode-score space, allowing adjustable degrees of freedom.

Main Results:

• Significantly reduced training costs due to decreased space complexity.
• Enabled projection of high-dimensional neural activity onto a low-dimensional mode space.
• Successfully applied the framework to digit classification and selective sensory integration tasks.

Conclusions:

• The Hopfield-like decomposition accelerates SNN training and reduces computational expense.
• The method reveals low-dimensional attractor structures within high-dimensional neural dynamics.
• This approach enhances both the efficiency and interpretability of SNNs.