TACOS: Task Agnostic Continual Learning in Spiking Neural Networks

TL;DR

This work shows that neuro-inspired mechanisms such as synaptic consolidation and metaplasticity can mitigate catastrophic interference in a spiking neural network, using only synapse-local information, with no need for task awareness, and with a fixed memory size that does not need to be increased when training on new tasks.

Abstract

Catastrophic interference, the loss of previously learned information when learning new information, remains a major challenge in machine learning. Since living organisms do not seem to suffer from this problem, researchers have taken inspiration from biology to improve memory retention in artificial intelligence systems. However, previous attempts to use bio-inspired mechanisms have typically resulted in systems that rely on task boundary information during training and/or explicit task identification during inference, information that is not available in real-world scenarios. Here, we show that neuro-inspired mechanisms such as synaptic consolidation and metaplasticity can mitigate catastrophic interference in a spiking neural network, using only synapse-local information, with no need for task awareness, and with a fixed memory size that does not need to be increased when training on new tasks. Our model, TACOS, combines neuromodulation with complex synaptic dynamics to enable new learning while protecting previous information. We evaluate TACOS on sequential image recognition tasks and demonstrate its effectiveness in reducing catastrophic interference. Our results show that TACOS outperforms existing regularization techniques in domain-incremental learning scenarios. We also report the results of an ablation study to elucidate the contribution of each neuro-inspired mechanism separately.

Figures (6)

• Figure 1: Supervised spiking neural network architecture described in this work. Error is propagated directly to all layers through random feedback, where the induction of plasticity is dependent on a Hebbian-like surrogate gradient rule. Unlike traditional neural networks, our model has more complex synapses which consist of the synaptic strength, a reference weight for heterosynaptic plasticity, and activity-dependent metaplasticity which regulates the degree of plasticity per synapse.
• Figure 2: Overview of key plasticity mechanisms for continual learning studied in spiking neural networks. Block 1 illustrates learning in a network without synaptic protection mechanisms. Block 2 (metaplasticity mechanisms) and block 3 (heterosynaptic plasticity) represent two of the new mechanisms incorporated in SNNs for this study.
• Figure 3: Accuracy of each task over time for TACOS with different maximum plasticity settings, a fixed metaplasticity model with different metaplasticity strengths, and a baseline SNN on split-MNIST. The dashed gray line indicates when the model has been trained on the task specified on the x-axis. Accuracy values to the left of the gray line reflect forward transfer induced from learning prior tasks, accuracy values to the right reflect backwards transfer from learning downstream tasks.
• Figure 4: We observe the plasticity in TACOS on split-MNIST by measuring; i) the average weight change induced after learning task 1 (blue), and ii) the average weight change during learning the final task (red). Left) Average change in input-hidden layer weights after learning each task. Right) Average change in hidden-output layer weights after learning each task.
• Figure 5: Continual learning on split-MNIST with reduced training sets (full test set used for accuracy). Blue accuracy bars represent TACOS with scaled metaplasticity and consolidation based on number of training samples, red accuracy bars reflect using the same parameters that worked best on the full dataset.