Voltage-dependent synaptic plasticity: Unsupervised probabilistic Hebbian plasticity rule based on neurons membrane...

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) are bio-inspired computational models.
• Implementing efficient learning rules on neuromorphic hardware is crucial for SNNs.
• Traditional synaptic plasticity rules like STDP have limitations in hardware implementation.

Purpose of the Study:

• To introduce a novel unsupervised, local learning rule called voltage-dependent-synaptic plasticity (VDSP).
• To enable online implementation of Hebb's plasticity mechanism on neuromorphic hardware.
• To reduce computational overhead and improve adaptability compared to existing methods.

Main Methods:

• Developed VDSP, a learning rule updating synaptic conductance based on postsynaptic neuron spikes and presynaptic neuron membrane potential.
• Performed mathematical analysis to establish equivalence between VDSP and STDP.
• Trained a single-layer SNN using VDSP for handwritten digit recognition on the MNIST dataset.

Main Results:

• VDSP reduces synaptic updates by half compared to STDP.
• Achieved 85.01% accuracy on MNIST with 100 neurons, improving to 90.56% with 500 neurons.
• VDSP demonstrated better adaptation to input signal frequency and robustness against hyperparameter tuning.

Conclusions:

• VDSP is a viable and efficient learning rule for SNNs on neuromorphic hardware.
• The proposed rule shows strong performance in spatial pattern recognition tasks.
• VDSP offers advantages in terms of computational efficiency and adaptability over STDP.