A Closer Look at Prototype Classifier for Few-shot Image Classification
This work addresses the need for efficient few-shot learning methods by providing a simpler, non-meta-learning approach that avoids retraining for new classes, though it is incremental as it builds on existing prototype classifier frameworks.
The paper tackles the problem of few-shot image classification by analyzing why prototype classifiers work well without meta-learning or retraining linear classifiers, finding that feature transformation via L2 normalization and minimizing within-class to between-class variance ratio achieves comparable performance to existing methods.
The prototypical network is a prototype classifier based on meta-learning and is widely used for few-shot learning because it classifies unseen examples by constructing class-specific prototypes without adjusting hyper-parameters during meta-testing. Interestingly, recent research has attracted a lot of attention, showing that training a new linear classifier, which does not use a meta-learning algorithm, performs comparably with the prototypical network. However, the training of a new linear classifier requires the retraining of the classifier every time a new class appears. In this paper, we analyze how a prototype classifier works equally well without training a new linear classifier or meta-learning. We experimentally find that directly using the feature vectors, which is extracted by using standard pre-trained models to construct a prototype classifier in meta-testing, does not perform as well as the prototypical network and training new linear classifiers on the feature vectors of pre-trained models. Thus, we derive a novel generalization bound for a prototypical classifier and show that the transformation of a feature vector can improve the performance of prototype classifiers. We experimentally investigate several normalization methods for minimizing the derived bound and find that the same performance can be obtained by using the L2 normalization and minimizing the ratio of the within-class variance to the between-class variance without training a new classifier or meta-learning.