Context Gating in Spiking Neural Networks: Achieving Lifelong Learning through Integration of Local and Global Plasticity

TL;DR

The paper tackles lifelong learning in spiking neural networks by introducing context gating (CG-SNN) that integrates local plasticity (STDP or Oja’s rule) with global backpropagation to emulate cognitive control. It presents two implementations, a single-spike and a multi-spike variant, and demonstrates through experiments with human-like cognitive tasks that CG-SNN can retain past knowledge while adapting to new tasks, with the multi-spike version showing superior memory retention and alignment with human data. The approach yields higher task-selectivity in neurons and reproduces key human behavioral patterns, such as the advantage of blocked over interleaved training and congruency effects, suggesting biological plausibility and potential for scalable neuromorphic deployment. The results indicate that combining local context gating with global learning offers a viable path to robust, energy-efficient lifelong learning in SNNs, with broad implications for neuro-inspired hardware and cognition modeling.

Abstract

Humans learn multiple tasks in succession with minimal mutual interference, through the context gating mechanism in the prefrontal cortex (PFC). The brain-inspired models of spiking neural networks (SNN) have drawn massive attention for their energy efficiency and biological plausibility. To overcome catastrophic forgetting when learning multiple tasks in sequence, current SNN models for lifelong learning focus on memory reserving or regularization-based modification, while lacking SNN to replicate human experimental behavior. Inspired by biological context-dependent gating mechanisms found in PFC, we propose SNN with context gating trained by the local plasticity rule (CG-SNN) for lifelong learning. The iterative training between global and local plasticity for task units is designed to strengthen the connections between task neurons and hidden neurons and preserve the multi-task relevant information. The experiments show that the proposed model is effective in maintaining the past learning experience and has better task-selectivity than other methods during lifelong learning. Our results provide new insights that the CG-SNN model can extend context gating with good scalability on different SNN architectures with different spike-firing mechanisms. Thus, our models have good potential for parallel implementation on neuromorphic hardware and model human's behavior.

Figures (6)

• Figure 1: The framework of the network construction of the single-spike CG-SNN and multi-spike CG-SNN. (left) The network structure of single-spike CG-SNN. The first encoding layer codes the input stimuli into spike trains and avoids weak signal ignorance. The additional context signals of tasks are introduced into the second layer. The iterative training between global and local plasticity proceeds during the training process, in which all the synaptic weights are updated by the global backpropagation rule, and the weights of context signal are updated by the local plasticity rule. (middle) The network structure of multi-spike CG-SNN. The network is a simple feed-forward MLP with two hidden layers with IF spiking neurons that receive the pixels information of pictures together with the one-hot contextual signal as input. And the connections between the contextual cue and the first layer will be successively updated by the Hebbian learning rule. (right) shows the blocked and interleaved distribution of the two tasks' samples.
• Figure 2: (A) The accuracy curve of single-spike CG-SNN. (left) Vanilla network of single-spike SNN. (mid) vanilla network with STDP learning rules. (right) vanilla network with multiple STDP update. (B) (left top) The target choice of the given data. (other) Plotting the choice of the trained network with (right top) vanilla setting, (left bottom) STDP learning rules, and (right bottom) multiple STDP updating. (C) (left top) The average absolute values of weights related to the context signal input and other weights in the hidden layer of three networks. (other) The time difference of spiking time of hidden layer neurons of a trained network under two different contextual signals. They are vanilla networks (right top), networks with STDP learning rules (left bottom), and networks with multiple STDP updating (right bottom).
• Figure 3: (top) The network accuracy change curve of the vanilla blocked training (left), applying OWM learning algorithm on the latter two connections and to all the connections (mid ), XdG method, and our method (right). (bottom) The standard reward value of the whole task and the choice of the trained network in two dimensions under the vanilla blocked training network (left two), the network with OWM applied in latter two layers (mid two), and our proposed network (right two).
• Figure 4: (A) (top) Test accuracy of human data and three network data: (left) human data, (mid left) vanilla network, (mid right) Flesh et al.'s model, (right) our model. (bottom) The comparison of congruency effect. (B)(top) Sigmoidal fits of choice of four conditions mentioned above. (bottom) Linear fitting of factorized and linear model to four conditions.
• Figure 5: Neuron's selectivity analysis. (left) Proportion of task selective neurons under blocked and interleaved training of the reference network (left) and our CG-SNN networks (right). (right) The average hidden layer neuron's activity of these two networks. The activation values and spiking frequency are computed for reference network and CG-SNN, respectively.