TLDR: A new framework called Multi-Plasticity Synergy Learning (MPSL) improves Spiking Neural Network (SNN) training by integrating multiple learning mechanisms like STBP, Hebbian, and SBP. Inspired by how the brain uses co-existing neurotransmitters, MPSL adaptively balances these mechanisms’ contributions, leading to significantly better accuracy and robustness on various datasets without increasing inference costs.
Spiking Neural Networks (SNNs) are gaining significant attention as brain-inspired models, celebrated for their energy efficiency and strong potential in processing temporal information. However, a major hurdle remains: effectively training these networks. Unlike the human brain, which employs a multitude of coexisting learning strategies, most current SNN training methods rely on a single form of synaptic plasticity, limiting their adaptability and ability to represent complex information.
Addressing this challenge, researchers Yuzhe Liu, Xin Deng, and Qiang Yu have introduced a novel training framework called Multi-Plasticity Synergy Learning (MPSL). This biologically inspired approach integrates multiple synergistic plasticity mechanisms to enhance SNN training, allowing diverse learning algorithms to work together while maintaining their unique update dynamics. The core idea draws inspiration from the biological phenomenon of neurotransmitter co-existence, where individual neurons can release multiple neurotransmitters that influence synaptic activity through distinct yet synchronized pathways.
The MPSL framework computationally emulates this biological principle. It enables different learning mechanisms, each with its own strengths, to collaboratively guide the accumulation of membrane potential within the SNN. For instance, the framework incorporates Spatio-temporal Backpropagation (STBP) for global, gradient-based optimization, Hebbian learning for local, correlation-driven adaptation, and Self-backpropagation (SBP) for local feedback and gradient approximation. A key innovation is the introduction of a learnable modulation parameter. This parameter adaptively balances the contribution of each mechanism during the fusion process, much like how the quantities of co-released neurotransmitters and their corresponding receptors vary across neuron types in the brain.
The interaction between these mechanisms is carefully designed. Global error signals from STBP not only update its own synaptic weights but also guide the optimization of parameters in Hebbian learning and SBP, providing a form of indirect supervision. Furthermore, SBP utilizes the synaptic update signals derived from Hebbian learning, reflecting a biologically grounded coordination where plasticity in one pathway informs another. Ultimately, all these mechanisms collectively contribute to the membrane potential dynamics, forming a coherent and synergistic learning process that enriches temporal dynamics and improves representational ability.
A significant advantage of the MPSL framework is its efficiency. While it uses dedicated weight components for each mechanism during training, these are linearly merged into a single effective weight for inference using the learned coefficients. This means the deployed model maintains the same parameter footprint as conventional single-rule models, introducing no extra computational cost at runtime.
The effectiveness and robustness of the MPSL method were rigorously evaluated on both static image datasets (MNIST, Fashion-MNIST, CIFAR-10) and dynamic neuromorphic datasets (N-MNIST, DVS-Gesture). The results consistently showed that MPSL achieves superior performance compared to conventional learning mechanism models. For example, it reached 99.52% accuracy on MNIST and 97.22% on DVS-Gesture, outperforming methods like STDP, SBP, STBP, and HGLL. Beyond accuracy, the framework demonstrated enhanced resilience to input perturbations, maintaining significantly higher accuracy under various levels of Gaussian noise, salt-and-pepper noise, and image cropping.
Also Read:
- Brain-Inspired Neural Network Design Boosts Efficiency of Equilibrium Propagation
- Boosting Spiking Neural Network Accuracy with Adaptive Neurons and Complexity Analysis
An ablation study further highlighted the importance of the learnable fusion coefficients, showing that dynamically tuning these parameters is crucial for optimal performance. Feature visualizations using t-SNE revealed that the full MPSL model produces more compact and clearly separated feature clusters, indicating improved class discrimination and representational capacity. This research provides a general and extensible foundation for developing more powerful SNNs, guided by multi-strategy brain-inspired learning. For more details, you can read the full research paper here.
