STDP-based spiking deep convolutional neural networks for object recognition.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spike-timing-dependent plasticity (STDP) enables unsupervised visual feature extraction in shallow spiking neural networks (SNNs).
• Deep learning with rate-based networks demonstrates improved recognition robustness through multi-layered architectures.
• Previous SNNs had limitations in depth and trainable layers, hindering complex feature learning.

Purpose of the Study:

• To design and evaluate a deep SNN architecture combining STDP with temporal coding for unsupervised visual learning.
• To investigate hierarchical feature extraction and recognition capabilities in deep SNNs.
• To compare the performance of STDP-based deep SNNs against other unsupervised methods.

Main Methods:

• Developed a deep SNN with multiple convolutional layers trained via STDP and pooling layers.
• Employed a temporal coding scheme where neuron firing latency indicates activation strength.
• Trained the network on natural images and evaluated performance on benchmark datasets (Caltech 101, ETH-80, MNIST).

Main Results:

• The deep SNN learned hierarchical features, from edges to object prototypes, using only a few examples per category without labels.
• Feature learning was efficient, with sparse coding (thousands of spikes per image) and potential category inference from single higher-order neurons.
• STDP-based deep SNNs outperformed random crops (HMAX) and auto-encoders in unsupervised learning tasks.

Conclusions:

• Combining STDP with latency coding in deep SNNs offers a promising approach for unsupervised visual learning.
• This approach mimics biological visual systems in terms of learning, speed, and energy efficiency.
• The developed deep SNNs have significant potential for artificial vision systems, especially in hardware implementations.