Representation learning using event-based STDP.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Traditional neural network representation learning is mature, but spiking neural network (SNN) models remain challenging.
• Developing efficient and biologically plausible learning rules for SNNs is crucial for advancing neuromorphic computing.

Purpose of the Study:

• To propose an event-based method for training a feedforward SNN layer for visual feature extraction.
• To introduce novel spike-timing-dependent plasticity (STDP) and threshold adjustment rules for SNNs.
• To achieve sparse and independent spiking representations.

Main Methods:

• Developed a novel STDP learning rule derived from a vector quantization objective with a sparsity constraint.
• Introduced a threshold adjustment rule for independence and sparsity.
• Utilized a softmax function for inhibition and implemented a form of spike-based competitive learning.
• Employed leaky, integrate-and-fire (LIF) neurons for the spiking network.

Main Results:

• Demonstrated a sparse spiking visual representation model.
• Achieved low reconstruction loss comparable to state-of-the-art visual coding approaches.
• Validated the method on MNIST and natural image datasets.

Conclusions:

• The proposed method enables biologically plausible and hardware-friendly local learning rules for SNNs.
• The event-based approach facilitates efficient visual feature extraction using sparse spiking representations.
• This work advances the development of neuromorphic computing systems.