Adaptive Reorganization of Neural Pathways for Continual Learning with Spiking Neural Networks

TL;DR

A brain-inspired continual learning algorithm with adaptive reorganization of neural pathways, which employs Self-Organizing Regulation networks to reorganize the single and limited Spiking Neural Network (SOR-SNN) into rich sparse neural pathways to efficiently cope with incremental tasks.

Abstract

The human brain can self-organize rich and diverse sparse neural pathways to incrementally master hundreds of cognitive tasks. However, most existing continual learning algorithms for deep artificial and spiking neural networks are unable to adequately auto-regulate the limited resources in the network, which leads to performance drop along with energy consumption rise as the increase of tasks. In this paper, we propose a brain-inspired continual learning algorithm with adaptive reorganization of neural pathways, which employs Self-Organizing Regulation networks to reorganize the single and limited Spiking Neural Network (SOR-SNN) into rich sparse neural pathways to efficiently cope with incremental tasks. The proposed model demonstrates consistent superiority in performance, energy consumption, and memory capacity on diverse continual learning tasks ranging from child-like simple to complex tasks, as well as on generalized CIFAR100 and ImageNet datasets. In particular, the SOR-SNN model excels at learning more complex tasks as well as more tasks, and is able to integrate the past learned knowledge with the information from the current task, showing the backward transfer ability to facilitate the old tasks. Meanwhile, the proposed model exhibits self-repairing ability to irreversible damage and for pruned networks, could automatically allocate new pathway from the retained network to recover memory for forgotten knowledge.

Figures (5)

• Figure 1: Sparse neural pathways self-organized collaboration for continual learning. Purple neurons and cyan neurons are individual neurons for task 1 and task 2, respectively, and blue neurons are shared for both tasks. In the blue box, the different synapses of neuron D are utilized for different tasks and form sparse connections.
• Figure 2: The procedure of SOR-SNN model. Each spiking neural network block in the proposed SOR-SNN model involves a self-organizing regulation network which is responsible for selectively activating task-specific sparse pathways in the SNN. For example: the purple connections form the pathway for task 1. In particular, the self-organizing regulatory network contains the fundamental weighing module and the path searching module. The large number of different combinations of connections allows the limited SNN to have the capacity to incrementally learn more n tasks.
• Figure 3: Validation of child-like simple-to-complex continual learning. (A-C) The simple to complex cognitive tasks include sketches, cartoons and photos. (E-G) Task-specific sparse pathways, for example, the blue, yellow and purple arrows represent the pathways for Task 1, Task 2 and Task 3 respectively in the fully connected output layer. (D,H,L) Visualization of synaptic activation counts in partial convolutional and fully connected layers. (I-K) Distribution of real-valued weights in the fully connected layer for three different tasks.
• Figure 4: The comparative performance of SOR-SNN on diverse continual learning tasks. The average accuracy (A-C) and the number of inactive parameters (D-F) of the network for simple to complex cognitive tasks, the CIFAR100 and Mini-ImageNet datasets. The average accuracy on the large scale ImageNet dataset (G).
• Figure 5: (A) The current test accuracy of past learned tasks in our SOR-SNN model. (B-C) The effect of memory loss coefficient and orthogonal loss coefficient on performance. (D) Injury schematic, containing the initial network, the network with task-specific pathways assigned and the network after injury task 1 of SNNs. (E) Accuracy comparisons before and after injury of first four tasks on CIFAR100 10steps.