FEW-SHOT LEARNING WITH REPRESENTATIVE GLOBAL PROTOTYPE

Abstract

Few-shot learning is often challenged by low generalization performance due to the assumption that the data distribution of novel classes and base classes is similar while the model is trained only on the base classes. To mitigate the above issues, we propose a few-shot learning with representative global prototype method. Specifically, to enhance the generalization to novel classes, we propose a method to jointly train the base classes and the novel classes, using selected representative and non-representative samples to optimize representative global prototypes, respectively. Additionally, a method that organically combines the sample of base classes conditional on semantic embedding to generate new samples of novel classes with the original data is proposed to enhance the data of novel classes. Results show that this training method improves the model's ability to describe novel classes, improving the classification performance for a few shots. Intensive experiments have been conducted on two popular benchmark datasets, and the experimental results show that this method significantly improves the classification ability of few-shot learning tasks and achieves state-of-the-art performance.

1. INTRODUCTION

The advantages of deep learning depend on big data, and machines can learn effectively under the drive of big data. However, an insufficient amount of data leads to problems, such as model overfitting, which makes the model not fit well on the dataset outside the training data, showing weak generalization ability Kang et al. (2019) . Thus, few-shot learning for small datasets has become a key technology to solve this type of problem. Few-shot learning (FSL) refers to the use of a small amount of labeled data per class training to make the model show higher performance, in line with the law of human learning. Meta-learning is the basis of most of the existing FSL methods. In meta-learning, training is modeled after a test, i.e., both base and novel classes learn a task consisting of a few samples from N-way K-shot, after which knowledge is transferred from the base classes in the form of appropriate initial conditions Finn et al. ( 2017 (2017) . There is, however, a fundamental limitation to all of these approaches: the models assumed that the distribution of base class data and novel class data was similar, so they mostly learned using only base class data, which does not guarantee that they are generalizable to novel class data. A few-shot learning method uses the transferability of knowledge to learn the ability to characterize samples from novel classes with only a few markers, while traditional machine learning and deep learning feed all classes of samples to the model at the beginning of training. Once the data distribution has low consistency, transferred knowledge is not applicable to generalize the novel class of data. Thus 2018) is used to enhance the samples of the novel classes to alleviate the sample imbalance. However, since the global prototype is optimized directly using all samples, it is not representative. Moreover, the sample generation method actually requires Therefore, we propose a new FSL method to solve the problem of poor generalization ability for fewshot data and sample imbalance to better describe the data of the novel class. The novel class and the selected representative and non-representative data from the base class are fed together into the network to jointly learn the representational ability of the data, and thus it is called a representative global prototype. In our view, the FSL model does not have to meet the barrier of similar distribution of the novel and base class data, and is better suited to identify novel classes when it is trained with novel class data. However, trying to learn the global prototype of the base and novel class requires overcoming the problem of data imbalance with sparse data in the novel class. To overcome the strong correlation between the base and novel class data distribution, we propose a new sample generation strategy. Specifically, as in Xu & Le (2022), more representative samples are selected to generate more representative features for the novel class using the CVAE Sohn et al. (2015) generative model conditional on semantic embedding. In the above operation, we obtain enhanced data strongly correlated with the novel class, with the base class as auxiliary information. Our main contributions are summarized as follows: (1) We propose a novel jointly training strategy for few-shot learning via representative and non-representative samples to break the assumption of meta-learning. (2) We propose a sample synthesis method to enhance novel classes of data. (3) Experiments show that our method exhibits state-of-the-art performance on both miniImageNet and tieredImageNet datasets. (2020) . The models of the second category are learning a good metric, the core of which is to learn effective discriminative features through a kernel function. Specifically, metric learning calculates the similarity between two samples by learning an embedding module and a metric module. The



); Fei et al. (2020); Zhang et al. (2020), embeddings Vinyals et al. (2016); Chopra et al. (2005); Hadsell et al. (2006); Sung et al. (2018) or optimization strategies Ravi & Larochelle

, learning the global representation of all class data together can alleviate the problem of model overfitting on the base classes like machine learning and deep learning. The novel class data is learned at the beginning, and the metric novel class data and the global representation strengthen the discriminative nature of the novel class samples Wang et al. (2018); Li et al. (2019). At the same time, the sample synthesis strategy of Wang et al. (

Figure 1: An illustration of representative global prototype. Combining representative samples from the base class, semantic embeddings, and novel class samples allows us to generate new samples, which are then used to train a representative global prototype.

FEW-SHOT LEARNING Few-shot learning comes into play when we only have very limited training samples. Most of the recent deep learning approaches rely on meta-learning or learning-to-learn Hochreiter et al. (2001) strategy, which improves the performance of the novel task by the provided data set and the meta knowledge extracted across tasks by a meta-learner. Specifically, the meta-learner transfers knowledge learned from many base class tasks to help few-shot learning to complete training tasks in novel classes. Presently, representative FSL via meta-learning can be divided into three categories: fine-tune based, metric based, and optimization based. (1) The methods of the first category aim to learn suitable initialization parameters for training novel class samples in order to train a novel classifier faster and better Finn et al. (2017); Rusu et al. (2019); Nichol & Schulman (2018); Fei et al. (2020); Zhang et al.

