MetaLearning

Few-shot classification aims to learn a classifier to recognize unseen classes during training with limited labeled examples.

• meta-learning paradigm: transferable knowledge is extracted and propagated from a collection of tasks to prevent over fitting and improve generalization.
  • model initialization based methods;
  • metric learning methods
  • hallucination based methods
• directly predicting the weighs of the classifiers for novel classes

Relative Work

• Initialization based methods:
  • good model initialization: to learn to fine-tune, learn with limited number of labeled examples and small number of gradient update steps;
  • learning an optimizer: LSTM-based meta-learner for replacing the stochastic gradient decent optimizer.
  • these methods have difficulty in handling domain shifts between base and novel classes.
• Distance metric learning based methods: learn to compare, make their prediction conditioned on distance or metric to few labeled instances during training process. MatchingNet ProtoNet RelationNet MAML
  • cosine similarity
  • Euclidean distance to class-mean representation
  • CNN-based relation module
  • ridge regression
  • graph neural network
  • simply reducing intra-class variation in a baseline methods using base class data
• Hallucination based methods: learn to augment, learns a generator from data in the base classes, and use the learned generator to hallucinated new novel class data for data augmentation.
  • transfer appearance variations exhibited in the base classes:
    • transfer variance in base classes to novel classes, use GAN to transfer style
  • directly integrate the generator into a meta-learning algorithm.
• Domain adaption: reduce the domain shifts between source and target domain, or novel tasks in a different domain. Dong&Xing

Limitations:

• discrepancy of the implementation details among multiple few-shot learning algorithms obscures the relative performance gain;
• the novel classes are sampled from the same dataset, lack of domain shift between the base and novel classes makes the evaluation scenarios unrealistic.

Solution:

• using a deep backbone shrinks the performance gap between different methods in the setting of limited domain differences between base and novel classes.
• by replacing the linear classifier with a distance-based classifier is surprisingly competitive to exiting methods;
• practical evaluation setting where there exists domain shift between base and novel classes.

Datasets&Scenarios

• Scenarios: generic object recognition, fine-grained image classification, cross-domain adaptation
• Dataset
  • mini-ImageNet: a subset of 100 classes,contains 600 images for each class
  • CUB-200-2011 dataset: contains 200 classes and 11,788 images in total

1. Introdce

Triplet loss

2. Base Model

Resource

• Chen, Wei-Yu, et al. “A closer look at few-shot classification.” arXiv preprint arXiv:1904.04232 (2019). [pdf]