CDNA-SNN: A New Spiking Neural Network for Pattern Classification Using Neuronal Assemblies.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) represent the third generation of neural networks, closely mimicking biological neurons.
• SNNs offer higher computational capacity and lower power consumption than traditional sigmoidal neural networks.
• Existing SNNs can be improved by incorporating principles from human brain neuronal assemblies.

Purpose of the Study:

• Introduce a novel SNN architecture, the class-dependent neuronal activation-based SNN (CDNA-SNN).
• Develop a new learning algorithm and training method for CDNA-SNN based on neuronal assemblies and spike-timing-dependent plasticity (STDP).
• Evaluate the performance and efficiency of CDNA-SNN against established SNN training methods.

Main Methods:

• Designed CDNA-SNN with learnable Class-Dependent Neuronal Activations (CDNAs) indicating neuron activity per class.
• Developed a learning algorithm to form class-specific neuronal assemblies based on CDNAs.
• Implemented a novel STDP variant controlled by neuronal assemblies and CDNAs for training.
• Utilized nested cross-validation (N-CV) for hyperparameter optimization on benchmark datasets.

Main Results:

• CDNA-SNN outperformed SWAT and SpikeProp on 3/5 UCI datasets and SRESN on 2/5 UCI datasets.
• The proposed SNN achieved top performance on Fashion MNIST and comparable results on MNIST and N-MNIST.
• CDNA-SNN demonstrated significant reductions in trainable parameters (1%-35%) compared to other SNNs.

Conclusions:

• CDNA-SNN offers a powerful and efficient approach to Spiking Neural Network design.
• The integration of brain-inspired concepts like neuronal assemblies enhances SNN performance.
• CDNA-SNN presents a promising alternative for low-power, high-performance neural network applications.