Biologically inspired spiking diffusion model with adaptive lateral selection mechanism.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Biological neural circuits utilize lateral connections for local processing and learning.
• Spiking neural networks (SNNs) process information via binary spikes, offering potential for efficient computation.
• Substructure selection networks are crucial for refining complex data representations.

Purpose of the Study:

• To develop a novel diffusion model integrating lateral connections with a substructure selection network using SNNs.
• To enhance model adaptability and expressivity through iterative refinement of the substructure selection network.
• To demonstrate biological plausibility and superior performance compared to existing SNN generative models.

Main Methods:

• Integration of lateral connections within a substructure selection network framework.
• Implementation of a lateral connection mechanism using spiking neurons, including a learnable lateral matrix and mapping function.
• Mathematical modeling to establish the biological plausibility of the lateral update mechanism based on synaptic plasticity principles.

Main Results:

• The proposed lateral update mechanism aligns with biologically plausible synaptic plasticity.
• Substructure selection and lateral connections play significant roles in model training and performance.
• The novel SNN-based diffusion model consistently outperforms state-of-the-art SNN generative models on benchmark datasets.

Conclusions:

• The developed SNN-based diffusion model with integrated lateral connections offers enhanced adaptability and expressivity.
• The approach demonstrates a biologically plausible mechanism for synaptic plasticity in artificial neural networks.
• This work advances the capabilities of SNNs for generative modeling tasks.