RRAM-Based Spiking Neural Network With Target-Modulated Spike-Timing-Dependent Plasticity.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Neuromorphic Engineering

Background:

• Spiking neural networks (SNNs) for image classification often require numerous neurons or external classifiers.
• Conventional bio-inspired learning rules like STDP have limitations in performance and learning opportunities.
• Existing methods struggle with efficiency and robustness in SNN-based image recognition.

Purpose of the Study:

• To introduce a novel bio-plausible learning rule, target-modulated STDP (TSTDP), for improved SNN training.
• To propose an SNN architecture optimized for TSTDP, enabling efficient, high-accuracy image classification.
• To eliminate the need for external classifiers in SNNs through enhanced intrinsic learning capabilities.

Main Methods:

• Development of the target-modulated STDP (TSTDP) learning rule.
• Design of a new SNN architecture compatible with TSTDP and temporal spike encoding.
• Training and evaluation of the proposed SNN on benchmark datasets: MNIST, CIFAR-10, and DVS gestures.

Main Results:

• Achieved 92% accuracy on MNIST, a 7% improvement over conventional SNNs of similar size.
• Demonstrated up to 75% reduction in network size for comparable accuracy on MNIST.
• Showcased significant accuracy improvements on CIFAR-10 (up to 12.4%) and DVS gestures (up to 3.6%).
• Highlighted enhanced resilience to process variations compared to conventional networks.

Conclusions:

• TSTDP and the proposed SNN architecture offer a more efficient and accurate approach to neuromorphic image classification.
• The new method significantly reduces hardware requirements and improves robustness, paving the way for practical SNN applications.
• This work advances bio-inspired learning rules and SNN design for superior performance in real-world tasks.