小样本元学习综述：A Concise Review of Recent Few-shot Meta-learning Methods

Posted 2020-12-23 joselynzhao

原文链接 小样本学习与智能前沿 。 在这个公众号后台回复‘CRR-FMM‘，即可获得电子资源。

1 Introduction

In this short communication, we present a concise review of recent representative meta- learning methods for few-shot image classification. We re- fer to such methods as few-shot meta-learning methods. Af- ter establishing necessary notation, we first mathematically formulate few-shot learning and offer a pseudo-coded algo- rithm for general few-shot training and evaluation. We then provide a taxonomy and a gentle review of recent few-shot meta-learning methods, to help researchers quickly grasp the state-of-the-art methods in this field. Finally we summa- rize some vital challenges to conclude this review with new prospects.

MindMap

完整的mindmap：https://note.youdao.com/ynoteshare1/index.html?id=3cf041f5616d34bb4677d03434531c92&type=note

2. The Framework of Few-shot Meta-learning

2.1. Notation and definitions

A classifier ?? is expected to correctly discriminate query images Q conditional on the small-size labeled support images S

A good learner should not only extract sufficient transferable knowledge among tasks but also fast adapt to novel tasks.

Hence, in general, a few-shot meta-learning algorithm usu- ally consists of two components, a meta-learner component and a task-specific learner component.

Definition 1. (Small-sample learning)

Definition 2. (Few-shot learning)

Definition 3. (Few-shot meta-learning)

What to share, how to share and when to share are three components at the heart of few-shot meta-learning.

For example, embedding layers are often shared in a rigid manner (e.g. [2]) in fine-tuning; parameters optimized on the base dataset can be regarded as a good initialization (e.g. [4, 5]), for fast further learning conditional on few labeled samples from novel classes; and auxiliary information also helps few-shot learning, e.g. attribute annotations related to images [30].

2.2. Training and evaluation of few-shot meta-learning

In addition to standard few-shot episodes defined by ?? -way ?? -shot, other episodes can also be used as long as they do not poison the evaluation in meta- validation or meta-testing

In this section, we give a general few-shot episodic train- ing/evaluation guide in Algorithm 1

For few-shot meta-learning, we can al- ways design a deep neural network ???? parametrized by ?? as the classifier: we denote it as ???? (?|(S^*), (D_{base})), where (S^*) is some support set.

Note that in our notation (S^*)can also be a support set on the base classes (C_{base}) or even the whole base dataset (D_{base}), corresponding to the cases of meta-training or pre- training, respectively.

For few-shot classification problems, e.g., ??-way ??- shot classification, the performance of a learning algorithm is measured by its averaged accuracy on the query sets of the tasks generated on the novel dataset (D_{novel}) (i.e., the 15th line of Algorithm 1).

The meta-learner component is to learn transferable prior knowledge from the base dataset (D_{base}).

The existing few-shot meta-learning methods can be categorized into four branches according to their technical characteristics：

• 1 learning an initialization,
• 2 generation of parameters,
• 3 learning an optimizer,
• 4 memory-based methods.

3. Methods of Few-shot Meta-learning

3.1. Learning an initialization

The under- lying rationale is that the task-specific parameters are close to this shared global initialization for all the tasks generated from (D_{base}) and (D_{novel}).

It can be interpreted and executed in the following two ways in recent few-shot meta-learning methods:

• To learn a global initialization conditional on the (giant) base dataset
• To fine-tune the trained parameters on the base dataset (D_{base}) via conditioning from few labeled images on the novel dataset (D_{novel}).

MAML: the algorithm seeks to update the task-specific parameters and the global initialization jointly in an iterative manner.
TAML:which aims to train an initial model that is unbiased to all tasks.
TAML achieves the state-of-the-art performance in 5-way 1-shot and 5-way 5-shot classification on the Onimiglot dataset.

Either of these baseline and Baseline++ neural networks can be decom-posed into two parts:

• a convolution embedding network
• a classifier network.

In [2],at the fine-tuning stage, they only keep the learned embedding part and set up a new classifier fit for the ?? -way ?? -shot problems on tasks generated from (D_{novel}).

a linear mapping layer and a softmax activation function. The standard baseline method only learns the parameters of the linear mapping at the fine-tuning stage; a modified base- line method replaces the linear mapping layer by a layer that computes cosine distance between each image’s deep representation and the learned parameters of the linear mapping layer (learned at the fine-tuning stage).

Tokmakov et al. [30] .The proposed neural net- work is firstly trained on the base dataset (D_{base}), and then fine-tuned with those additional regularization terms condi- tional on the labeled images from each novel task.

3.2. Generation of parameters

The second branch focuses on rapid generation of parameters of task-specific neural networks from a meta-learner.

Munkhdalai and Yu [16]： The meta-learner is used to perform fast parameter generation for itself and the base-learner by minimizing both the representation loss and task loss across various tasks with an attention mechanism.

Fast parameter generation is achieved by MetaNet through learning the distribution of func- tional parameters of task-specific Matching Networks condi- tional on support sets of tasks.

Gidaris and Komodakis [6]：This is achieved by combining an attention- based classification weight generator and a cosine-based con- volution classifier which allows to learn both base and novel classes even at the testing stage.

Ren et al. [21] ：The whole training includes two stages.
A pre-training stage is to learn a good representation and so- called slow weights (??????????) of the top fully connected layer of the classifier. Then, an incremental few-shot episodic training is designed to increment novel classes into training via an episodic style.

On Mini-ImageNet, the MTL achieves the state-of-the-art per- formance, for 5-way 1-shot classification.

3.3. Learning an optimizer

Ravi and Larochelle [20] ：Their main contribution is to represent parameter optimization of a task-specific classifier by the evolution of LSTM’s cell states. Their work also uses a standard episodic meta training/evaluation as in Algorithm 1.

3.4. Memory-based methods

Mishra et al. [15] proposed a class of generic meta-learner architectures, called simple neural attentive learner (SNAIL), that combined temporal convolu- tions and soft attention for leveraging information from past episodes and for pinpointing specific pieces of information, respectively.

Munkhdalai et al. [17] proposed a neural mechanism, conditional shifted neurons (CSNs), which was capable of extracting conditioning information and producing condi- tional shifts for prediction in the process of meta-learning, and could be further incorporated into CNNs and RNNs.

4. Some Remaining Challenges

The main challenge of few-shot learning is the deficiency of samples.
Parameter-generation based meth- ods solve this problem by directly generating the parame-ters of the task-specific learner to mitigate the difficulty of training on novel data.

A better and more diversified meta-learner.

when a few- shot meta-learning algorithm has uneven performance on a series of tasks, the knowledge learned by the meta-learner can lead to large uncertainty in performance for novel tasks from unseen classes [10].
Apart from the existing few- shot meta-learning methods, meta-learning methods with di- versified emphases, such as learning a suitable loss function or learning a network structure, will also be valuable to ex- plore.

A more effective task-specific learner.

it remains vital for us to develop a feature extractor for task- specific learners that learns more discriminative features from only one or few labeled images.
Thus, it is important that task-specific learners are built on the loss functions that can ensure the robustness and performance of models.

Cross-domain few-shot meta-learning.

Inpractice,(D_{base}) and (D_{novel}) can be from different domains; such classification problems demand cross-domain few-shot learners.
Therefore, it merits further exploration on cross-domain few-shot meta-learning.

Multi-domain few-shot meta-learning.

If a meta-leaner can learn transferable knowledge from (D_{base}) consisting of multi-domain data, the meta-learner will be expected to have better gener- alization ability.

Reference

[2] Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B., 2019. A closer look at few-shot classification, in: International Conference on Learning Representations.
[15] Mishra, N., Rohaninejad, M., Chen, X., Abbeel, P., 2018. A simple neural attentive meta-learner, in: International Conference on Learn- ing Representations.
[17] Munkhdalai, T., Yuan, X., Mehri, S., Trischler, A., 2018. Rapid adap- tation with conditionally shifted neurons, in: International Confer- ence on Machine Learning, pp. 3661–3670.
[20] Ravi, S., Larochelle, H., 2017. Optimization as a model for few-shot learning, in: International Conference on Learning Representations.
[27] Sun,Q.,Liu,Y.,Chua,T.S.,Schiele,B.,2019.Meta-transferlearning
for few-shot learning, in: IEEE Conference on Computer Vision and
Pattern Recognition, pp. 403–412.