Deep CovDenseSNN: A hierarchical event-driven dynamic framework with spiking neurons in noisy environment

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Neurons utilize spike events to encode temporal information for neural computation.
• Spiking neural networks (SNNs) offer biologically relevant modeling but face limitations in sensory encoding and integration due to shallow structures and learning algorithms.

Purpose of the Study:

• To propose a novel hybrid framework, deep CovDenseSNN, integrating CNNs and SNNs.
• To enhance feature extraction and transmission in SNNs for improved performance, particularly in noisy environments.
• To provide a biologically realistic and adaptable model for image classification and neuromorphic hardware.

Main Methods:

• Developed a hybrid deep CovDenseSNN architecture combining CNN feature learning with SNN temporal processing.
• Utilized unsupervised learning rules within the SNN component for feature processing.
• Evaluated the model on MNIST and its variations to assess information extraction, transmission, and anti-noise abilities.

Main Results:

• The deep CovDenseSNN model demonstrated superior extraction and transmission of important information compared to existing models.
• The framework exhibited enhanced anti-noise ability in noisy environments.
• The architecture facilitates efficient feature representation and recognition within a temporal learning framework.

Conclusions:

• The proposed deep CovDenseSNN offers an effective approach to combine CNNs and SNNs for advanced image classification.
• The model improves information processing and robustness against noise, aligning with biological realism.
• This framework is suitable for neuromorphic hardware and provides insights into neural information transmission and representation.