Hierarchical Spiking-Based Model for Efficient Image Classification With Enhanced Feature Extraction and Encoding.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neuromorphic engineering

Background:

• Spiking neural networks (SNNs) offer computation efficiency and temporal encoding but struggle with spatio-temporal complexity and training optimization.
• Existing cognitive models face limitations in processing complex visual information and noise resistance.

Purpose of the Study:

• To propose a novel hierarchical event-driven visual device inspired by biological mechanisms for retina information processing.
• To integrate the functional learning of CNNs with the cognitive capabilities of SNNs in a biologically realistic framework.
• To enhance the efficiency and accuracy of visual data processing in artificial systems.

Main Methods:

• Developed a hybrid CNN-SNN framework incorporating biologically plausible unsupervised learning rules and advanced spike firing rate encoding.
• Modeled information transmission and representation within the retina using event-driven mechanisms.
• Implemented a novel quantization approach for efficient neuromorphic hardware deployment.

Main Results:

• The proposed model successfully processed greater vital data compared to existing cognitive models on MNIST and CIFAR-10 datasets, including noisy versions.
• Demonstrated superior focus accuracy and enhanced generalization ability through the joint CNN-SNN architecture.
• The quantization approach improved the model's efficiency for neuromorphic hardware implementation.

Conclusions:

• The hybrid CNN-SNN model offers a promising approach for efficient and accurate spatio-temporal data processing, inspired by biological vision systems.
• The biologically realistic design and advanced encoding methods contribute to improved performance and noise resistance.
• The developed model and quantization technique pave the way for more efficient neuromorphic computing applications.