Exploring Meta Learning: Parameterizing the Learning-to-learn Process for Image Classification

Meta-learning has emerged as a new paradigm in AI to challenge the limitation of conventional deep learning to acquire only task-specific knowledge. Meta-learning transcends this limitation by extracting the general concepts when learning tasks to apply these concepts later when learning new tasks. One popular meta-learning approach is model-agnostic meta-learning (MAML) which learns tasks by optimizing parameters towards highest generalizability of future tasks. The present paper applied a practical implementation of MAML to conduct an image classification task. Results showed that performance on learning new tasks neared training performance without overfitting. Furthermore, optimal values for inner-loop and outer-loop learning rate were close to default parameter values. Smaller batch sizes with more epochs improved learning in earlier epochs compared to larger batch sizes with fewer epochs. These findings show that MAML is able to transfer the concepts extracted during training effectively on to new tasks which it had not been trained on, similarly to how humans transfer knowledge.