Refined Meta-Learning Approach Leveraging Residual Networks for Enhanced Performance
Yihan Wang, Bo Wu, Hongchen Guo · 2023
Meta-learning, as a new way of small-sample learning, corrects the network parameters of the overall model for classification problems by treating tasks as inputs, thus obtaining a more rapid iteration effect and more generalizable network parameters. For current image classification problems, after model training, learning for new tasks often requires retraining with a large amount of new and past data, requiring the entire model parameters to be updated. Using the meta-learning method, the overall model parameters can be updated through sub-tasks, so as to quickly update the parameters according to new tasks, reducing the required training time and data. We propose an improved model based on the MAML algorithm, which has the characteristic of continuously updating the parameters through sub-tasks to obtain better initial training parameters, and focuses on obtaining higher accuracy on new tasks while ensuring training speed. This balances the accuracy and iteration speed of the model, and experimental results show that meta-learning can achieve a good balance between convergence speed and accuracy on new samples.