NG-SNN: A neurogenesis-inspired dynamic adaptive framework for efficient spike classification.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer low-power computation but face limitations in classifier accuracy and training efficiency.
• Hybrid SNN models decouple feature extraction and classification, concentrating computational load on the classifier.
• Fixed network topologies and costly surrogate-gradient training hinder SNN performance and adaptability.

Purpose of the Study:

• To develop a novel Spiking Neural Network architecture inspired by biological neurogenesis.
• To address the limitations of fixed topologies and computationally expensive training in current SNNs.
• To create a dynamic, adaptive framework for efficient and accurate SNN-based classification.

Main Methods:

• Introduced a neurogenesis-inspired spiking neural network (NG-SNN) with dynamic structural adaptation.
• Implemented a supervised incremental construction mechanism for task-optimal neuron integration.
• Developed an activity-dependent analytical learning method for efficient, single-shot weight computation.

Main Results:

• NG-SNN demonstrated dynamic structural adaptation and efficient non-iterative learning.
• The neurogenesis-driven approach resulted in a compact network structure with significantly fewer parameters.
• NG-SNN matched or surpassed competitor performance on diverse datasets without iterative training or manual tuning.

Conclusions:

• NG-SNN uniquely integrates dynamic structure and efficient learning for self-organizing, rapidly converging classification.
• The proposed model overcomes accuracy and efficiency bottlenecks in conventional SNN classifiers.
• NG-SNN offers a biologically plausible and computationally advantageous approach to neuromorphic computing.