few_shot_learning
few_shot_learning copied to clipboard
Awesome papers in few-shot learning/one-shot learning.
There are some papers/articles need to read.
Keywords:
- Meta learning
- Metric learning
- Few-shot learning
- One-shot learning
- Zero-shot learning
- GAN
- VAE
Metric learning based approaches
Papers:
These are the methods based on metric distance for few-shot learning.
Siamese network/Triplet network
- [x] Siamese Neural Networks for One-Shot Image Recognition
- [x] FaceNet: A Unified Embedding for Face Recognition and Clustering
- [x] Deep Metric Learning Using Triplet Network
Matching network
Some papers from Google DeepMind, should read in the order.
- [x] Neural Turing Machines
- [x] One-shot Learning with Memory-Augmented Neural Networks
- [x] Matching Networks for One Shot Learning
Prototypical network
Graph network
Papers:
Related articles:
Relation Network
Others
- [ ] Feature Generating Networks for Zero-ShsoetenLearning
- [x] Large Margin Few-Shot Learning
- [ ] Label-Embedding for Image Classification
- [ ] Evaluation of Output Embeddings for Fine-Grained Image Classification
- [ ] Learning Deep Representations of Fine-Grained Visual Descriptions
- [ ] Transductive Unbiased Embedding for Zero-Shot Learning
- Related articles: From Zero to Hero: Shaking Up the Field of Zero-shot Learning
- [ ] Recent Advances in Zero-shot Recognition
- [ ] An embarrassingly simple approach to zero-shot learning
Articles
Meta-Learning(learning to learn) based approaches
These are the methods based on meta learning for few-shot learning.
The most popular two are MAML and Reptile.
Papers
- [x] On First-Order Meta-Learning Algorithms
- [x] Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
- [x] META-LEARNING FOR SEMI-SUPERVISED FEW-SHOT CLASSIFICATION
- [x] Meta-Learning with Latent Embedding Optimization
Articles
This is the overview of learning to learn written by berkeley AI research.
This article introduced the meta-learning by animation clearly, one of the best explaination of meta-learning I think.
This article explain the Reptile, which is from the paper: On First-Order Meta-Learning Algorithms.
Generative and augmentation-based approaches
It's a new paper from NIPS 2018, by IBM research AI.
- [x] ∆-encoder: an effective sample synthesis method for few-shot object recognition
- [x] MetaGAN: An Adversarial Approach to Few-Shot Learning
- It's better to read the paper about MAML, Relation Network, DAGAN before.
- [ ] Data Agumentation Generative Adversarial Networks