Synthetic Meta-Learning: : Learning to learn real-world tasks with synthetic data

Abstract: Meta-learning is an approach to machine learning that teaches models how to learn new tasks with only a handful of examples. However, meta-learning requires a large labeled dataset during its initial meta-learning phase, which restricts what domains meta-learning can be used in. This thesis investigates if this labeled dataset can be replaced with a synthetic dataset without a loss in performance. The approach has been tested on the task of military vehicle classification. The results show that for few-shot classification tasks, models trained with synthetic data can come close to the performance of models trained with real-world data. The results also show that adjustments to the data-generation process, such as light randomization, can have a significant effect on performance, suggesting that fine-tuning to the generation process could further improve performance.