Few-Shot Learning with Feature Pairing and Mean Discrepancy

Abstract

Few-Shot Learning (FSL) is a sub-area of machine learning which mainly deals with data where there is a scarcity of training supervised samples. Few shot learning (FSL) more closely resembles the human brain in comparing new concepts to others based on prior experience rather than identifying it exactly. FSL aims to generalize the model across the tasks (in meta learning) opposed to the classical supervised learning which generalizes across the data points. In general the FSL models may suffer from underfitting because of scarcity of supervised samples and at the same time it causes overfitting as it is likely to memorize task specific features of the training set. This work aims to reduce such problems and is presented as a metric based model ”Few Shot Learning with Feature Pairing and Mean Discrepancy” (FL-FPMD). As the title suggests, feature pairing is one among various data augmentations. It is observed that flip augmentation is more suitable in the context of pairing the features within the given task. Memorizing task specific features is reduced by incorporating the discrepancy of mean distributions of the query and the support embedding in the loss function. The training and the evaluation is performed at the miniImageNet dataset and the results indicate that the proposed model outperforms the state-of-the-art models of similar complexity.