A Retinotopic Spiking Neural Network System for Accurate Recognition of Moving Objects Using NeuCube and Dynamic...

Area of Science:

• Computer Vision
• Neuroscience
• Artificial Intelligence

Background:

• Dynamic vision sensors (DVS) mimic the human retina, processing visual information via event-based spike trains.
• Spiking neural networks (SNNs) offer a biologically plausible model for brain-like computation.

Purpose of the Study:

• Introduce a new system for dynamic visual recognition.
• Combine bio-inspired hardware with a brain-like spiking neural network (NeuCube).
• Enhance understanding of neural network learning processes.

Main Methods:

• Utilize a dynamic vision sensor (DVS) to generate spike trains.
• Convolute spike trains and feed them into the 3D-organized NeuCube spiking neural network.
• Employ unsupervised learning with spike-timing-dependent plasticity, followed by supervised learning for classification.

Main Results:

• The system successfully learned spatio-temporal patterns from DVS data.
• Achieved 92.90% classification accuracy on the MNIST-DVS dataset.
• The NeuCube architecture allowed visualization of network connectivity during learning.

Conclusions:

• The proposed method shows promise for dynamic computer vision.
• Further exploration on diverse datasets is warranted.
• Potential for integration into multimodal systems for biologically plausible information processing.