Aggn: Augmented Gradient Guided Network for Few-Shot Regression

Abstract

Few-shot learning has become a pivotal focus in machine learning research due to its ability to quickly learn from minimal examples. This capability is highly valued, especially in applications where data collection is costly and time-consuming. Most research works in this field has concentrated on classification problems, but there has been much less focus on regression tasks. In this paper, we propose a novel approach (AGGN) for few-shot regression based on a differentiable reference function and a data augmentation strategy enhanced by the gradient-guided method. We use the reference function to guide learning by ensuring the model aligns with the same gradient information. Additionally, we employ the reference model in our data augmentation strategy to generate additional training data, leveraging the gradient information from the reference function to enhance the accuracy of new data values. Subsequently, the model’s parameters will be updated by combining real data, augmented data, and the differentiable reference function. We compare our model and state-of-the-art models, and the results confirm the superior performance and robustness of this approach.

Keywords: Few-shot learning, regression, Differentiable reference model, Transfer Learning, data augmentation