Tuning Convolutional Spiking Neural Network With Biologically Plausible Reward Propagation

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) mimic biological neurons, offering robust computation at low cost but are challenging to train with standard backpropagation (BP).
• The non-differentiable nature and event-based dynamics of SNN neurons hinder direct application of BP, which is also biologically implausible.
• Existing SNN training methods often lack biological realism, limiting our understanding of brain-inspired intelligence.

Purpose of the Study:

• To introduce a biologically plausible reward propagation (BRP) algorithm for training SNNs.
• To demonstrate the effectiveness of BRP in SNNs incorporating spiking-convolutional and full-connection layers.
• To compare the performance and computational efficiency of BRP-trained SNNs against traditional ANNs and BP-based SNNs.

Main Methods:

• Proposed a novel biologically plausible reward propagation (BRP) algorithm.
• Applied BRP to SNN architectures featuring 1-D and 2-D spiking convolutional kernels and full-connection layers.
• Utilized target labels propagated from the output to hidden layers, inducing local synaptic modifications via pseudo-BP or STDP.

Main Results:

• BRP-trained SNNs achieved accuracy comparable to state-of-the-art BP-based SNNs on spatial (MNIST, Cifar-10) and temporal (TIDigits, DvsGesture) tasks.
• SNNs trained with BRP demonstrated a 50% reduction in computational cost compared to traditional ANNs.
• The BRP algorithm showed consistency with top-down reward-guiding learning principles observed in the neocortex.

Conclusions:

• The BRP algorithm offers a biologically plausible and computationally efficient method for training SNNs.
• This approach bridges the gap between biologically realistic SNNs and effective training methodologies.
• Further research into biologically plausible learning rules for SNNs can enhance our understanding of biological intelligence.