Effective Active Learning Method for Spiking Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Deep Spiking Neural Networks (SNNs) demand substantial labeled data for high performance, which is costly to acquire.
• Conventional active learning methods are less effective in SNNs compared to Artificial Neural Networks (ANNs) due to differing feature representations and information transmission.

Purpose of the Study:

• To develop an effective active learning strategy tailored for deep SNN models.
• To address the limitations of existing active learning techniques in the context of SNNs.

Main Methods:

• Proposed a novel loss prediction module, ActiveLossNet, to extract features and identify valuable data samples for SNNs.
• Derived a corresponding active learning algorithm specifically designed for deep SNN architectures.
• Conducted experiments on diverse datasets (CIFAR-10, MNIST, Fashion-MNIST, SVHN) using various SNN frameworks (CIFARNet, ResNet-18).

Main Results:

• The proposed active learning algorithm significantly outperformed random selection and conventional ANN-based active learning methods.
• The method demonstrated faster convergence compared to existing active learning strategies.
• Achieved superior performance in deep SNN model training with reduced labeled data requirements.

Conclusions:

• The ActiveLossNet approach offers an effective solution for reducing data annotation costs in deep SNN training.
• This method enhances the efficiency and performance of active learning for SNNs, paving the way for more practical applications.
• The proposed algorithm represents a significant advancement in making SNNs more accessible and efficient for complex tasks.