Meta-Learning with Test-Time-Train
Tianao Wang · 2024
Meta-learning algorithm aim to distill experiences from past tasks to achieve rapid adaptation to new tasks. This approach has seen significant success in recent years. However, meta-learning still performs poorly when facing cross-domain tasks, primarily due to the training domain's task set not always providing effective experiences for the test domain's task set. To address this issue, we propose the Meta-learning with Test-time-train algorithm, which combine Test-time-train technique with meta-learning algorithm to minimize training loss of mixed data, enabling the model to adapt more quickly to unseen task distribution. For the training phase, we advocate for using mixed data from trainset and validation set to fine-tune parameter within the outer loop, allowing the model to proactively adapt to different task feature. For the testing phase, we advocate for using mixed data from validation set and test set to fine-tune parameter within the inner loop, making better prediction for unseen tasks. This method constitutes a direct meta-learning framework that seamlessly integrates with existing meta-learning methods to enhance their performance. Experimental evaluation confirm the effectiveness of the proposed framework, demonstrating state-of-the-art performance across different datasets.