Similarity-based context aware continual learning for spiking neural networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Biological brains adaptively coordinate neuronal populations for continuous learning in dynamic environments.
• Current spiking neural network (SNN) continual learning algorithms lack efficiency due to uniform task treatment and limited knowledge utilization.

Purpose of the Study:

• To propose a novel Similarity-based Context Aware Spiking Neural Network (SCA-SNN) algorithm for efficient continual learning.
• To enhance knowledge utilization and reduce energy consumption in incremental and class-incremental learning scenarios.

Main Methods:

• Developed the SCA-SNN algorithm inspired by brain's context-dependent plasticity.
• Implemented adaptive neuron reuse and flexible neuron expansion based on inter-task contextual similarity.
• Evaluated performance on diverse datasets including CIFAR100, ImageNet, FMNIST-MNIST, and SVHN-CIFAR100.

Main Results:

• SCA-SNN demonstrated superior performance over existing SNN and deep neural network (DNN) based continual learning algorithms.
• The model achieved efficient task incremental learning and class incremental learning.
• Adaptive neuron selection for related tasks enhanced biological interpretability.

Conclusions:

• The SCA-SNN algorithm significantly improves knowledge utilization and reduces energy consumption in continual learning.
• This approach offers a biologically plausible and efficient method for continual learning in artificial neural networks.
• SCA-SNN provides a promising direction for developing more adaptive and interpretable AI systems.