A Continual Learning Method for Facial Recognition Task

Most existing continual learning algorithms are applied to easy recognition tasks with smaller spans. In view of this, this paper attempts to apply continual learning algorithm to face recognition tasks with larger spans and higher difficulty, and analyze the performance of the algorithms on this type of task. In order to meet the demand for lightweight models in production, a new continual learning strategy involving the SqueezeNet, which has the same accuracy as AlexNet but 50 times smaller model parameters, with the Elastic Weight Consolidation (EWC), that is a regularization continual learning method with the least impact on model size is proposed. Based on this, the paper further explores the relationship between the difficulty of learning tasks and the learning order within the framework of continual learning. The experiments show that under the learning order of learning difficult tasks first and then learning easy tasks, the proposed method retained $16.1 \%$ and $11 \%$ more knowledge than typical deep learning models. Therefore, the proposed model is more suitable for facial recognition scenarios where difficult tasks are learned first and easy tasks are learned later, and can effectively avoid catastrophic forgetting problems in this scenario.