meta-transfer learning for few-shot learningyaliu/files/meta-transfer... · 2020-07-14 ·...

1

Meta-Transfer Learning for Few-Shot Learning Qianru Sun 1,3* Yaoyao Liu 1,2* Tat-Seng Chua 1 Bernt Schiele 3 1 National University of Singapore, 2 Tianjin University, 3 Max Planck Institute for Informatics Motivation & Contributions Meta-Transfer Learning Hard Task Meta-Batch ● Few-shot learning is challenging due to the lack of training data. ● Re-thinking promising methods: ○ deep neural networks (DNN) ○ transfer learning (pre-train, fine-tune) ○ meta-learning (meta gradient descent) ● Problem: “catastrophic forgetting” ● Our solution: Scaling & Shifting (SS) ● Problem: slow meta-training convergence ● Our solution: hard negative task sampling Detail Top performance is achieved! Faster convergence is achieved! Detail ● miniImageNet dataset Few-shot CIFAR-100 (FC100) dataset method backbone 1-shot 5-shot ● Few-shot CIFAR-100 (FC100) dataset method backbone 1-shot 5-shot ★ Code is available at: https://github.com/y2l/ meta-transfer-learning -tensorflow

Upload: others

Post on 19-Jul-2020

4 views

Category:

Documents

0 download

Report

Download

Embed Size (px):

TRANSCRIPT

Page 1: Meta-Transfer Learning for Few-Shot Learningyaliu/files/meta-transfer... · 2020-07-14 · Meta-Transfer Learning for Few-Shot Learning Qianru Sun1,3* Yaoyao Liu1,2* Tat-Seng Chua1

Meta-Transfer Learning for Few-Shot LearningQianru Sun1,3* Yaoyao Liu1,2* Tat-Seng Chua1 Bernt Schiele3

1National University of Singapore, 2Tianjin University, 3Max Planck Institute for Informatics

Motivation & Contributions Meta-Transfer Learning Hard Task Meta-Batch● Few-shot learning is challenging

due to the lack of training data.● Re-thinking promising methods:○ deep neural networks (DNN)○ transfer learning (pre-train, fine-tune)○ meta-learning (meta gradient descent)

● Problem: “catastrophic forgetting”

● Our solution: Scaling & Shifting (SS)

● Problem: slow meta-training convergence

● Our solution: hard negative task sampling

Detail

Top performance is achieved!

Faster convergence is achieved!

Detail

● miniImageNet dataset

Few-shot CIFAR-100 (FC100) dataset

method backbone 1-shot 5-shot

● Few-shot CIFAR-100 (FC100) datasetmethod backbone 1-shot 5-shot

★ Code is available at: https://github.com/y2l/meta-transfer-learning-tensorflow

Meta-transfer Learning for Few-shot Learningyaliu/files/meta... · Datasets miniImageNet - Reorganized from ImageNet - Vinyals et al.[1] first devised the dataset, and it is widely

Few-shot Slot Tagging with Collapsed Dependency Transfer

Zero shot learning through cross-modal transfer

Emergent Complexity and Zero-shot Transfer via

One-Shot Visual Imitation Learning via Meta-Learningproceedings.mlr.press/v78/finn17a/finn17a.pdf · Keywords: imitation learning, deep learning, one-shot learning, meta-learning

Meta-Transfer Learning for Few-Shot Learning€¦ · into two classes: data augmentation and task-based meta-learning. Data augmentation is a classic technique to in-crease the amount

LSTD: A Low-Shot Transfer Detector for Object …Figure 1: Basic deep architecture of low-shot transfer detector (LSTD). Since we aims at detecting objects effectively in the low-shot

UniT: Uniﬁed Knowledge Transfer for Any-shot Object

Few-shot Network Anomaly Detection via Cross-network Meta

Few-shot Knowledge Transfer for Fine-grained Cartoon Face

LSTD: A Low-Shot Transfer Detector for Object Detection

SML: Semantic Meta-learning for Few-shot Semantic ...SML: Semantic Meta-learning for Few-shot Semantic Segmentation Ayyappa Kumar Pambala1, Titir Dutta1, Soma Biswas1 1 Electrical

Meta-Transfer Learning for Zero-Shot Super-Resolutionopenaccess.thecvf.com/content_CVPR_2020/papers/Soh_Meta... · 2020. 6. 28. · Meta-Transfer Learning for Zero-Shot Super-Resolution

DARLA: Improving Zero-Shot Transfer in Reinforcement Learning · vironments (Jaco arm, DeepMind Lab) and base ... DARLA: Improving Zero-Shot Transfer in Reinforcement Learning of

Zero-shot Knowledge Transfer via Adversarial Belief Matching · Zero-shot Knowledge Transfer via Adversarial Belief Matching Paul Micaelli University of Edinburgh {paul.micaelli}@ed.ac.uk

Policy Transfer Algorithms for Meta Inverse Reinforcement ...Policy Transfer Algorithms for Meta Inverse Reinforcement Learning Benjamin Kha Abstract Inverse reinforcement learning

Meta-Transfer Learning for Few-Shot Learningopenaccess.thecvf.com/content_CVPR_2019/papers/Sun_Meta...Meta-Transfer Learning for Few-Shot Learning Qianru Sun1,3∗ Yaoyao Liu2∗ Tat-Seng

OOD-MAML: Meta-Learning for Few-Shot Out-of-Distribution … · In meta-testing phase, given K-shot N-way samples, we construct Nsub-tasks of K-shot OOD detection with respect to

One-shot Transfer Learning for Population Mapping

Meta-Reinforced Synthetic Data for One-Shot Fine-Grained ... · Meta-Reinforced Synthetic Data for One-Shot Fine-Grained Visual Recognition Satoshi Tsutsui ... three classiﬁers

MEDA: Meta-Learning with Data Augmentation for Few-Shot

Few-Shot Learning With Embedded Class Models and Shot-Free ... · Few-Shot Learning with Embedded Class Models and Shot-Free Meta Training Avinash Ravichandran Amazon Web Services

Shape-From-Recognition: Recognition enables Meta-data Transfer › info-stuff › publications › CVIU... · Shape-From-Recognition: Recognition enables Meta-data Transfer Alexander

Meta-Reinforced Synthetic Data for One-Shot Fine-Grained ...papers.nips.cc/paper/8570-meta-reinforced... · Few-shot Meta-learning. Few shot classiﬁcation [4] is a sub-ﬁeld of

One-shot Learning(2017-2018)ice.dlut.edu.cn/valse2018/ppt/07.one_shot_add_v2.pdf · •META-LEARNING FOR SEMI-SUPERVISED FEW-SHOT CLASSIFICATION (ICLR 2018) •One-shot Learning with

Few-Shot Human Motion Prediction via Meta-learning · Few-Shot Human Motion Prediction via Meta-learning Liang-Yan Gui(B), Yu-Xiong Wang, Deva Ramanan, and Jos´eM.F.Moura Carnegie

Meta-Learning: from Few-Shot Learning to Rapid

Meta-Learning for Few-Shot Land Cover Classification...The meta-learner learns how to learn by training and evaluating on the meta-training set. Meta-learning hyperparameters are tuned

ONE-SHOT PARAMETRIC AUDIO PRODUCTION STYLE TRANSFER …njb/research/icassp2020_style_transfer.… · ONE-SHOT PARAMETRIC AUDIO PRODUCTION STYLE TRANSFER WITH APPLICATION TO FREQUENCY

Zero-Shot Autonomous Vehicle Policy Transfer: From

Avatar-Net: Multi-Scale Zero-Shot Style Transfer by

Meta-Transfer Learning for Zero-Shot Super-Resolution...We found that simply employing transfer learning or ﬁne-tuning from a pre-trained network does not yield plausible results

Few-Shot One-Class Classiﬁcation via Meta-Learning

Learn to Cross-lingual Transfer with Meta Graph Learning ...Meta learning, a.k.a learning to learn, (Finn et al.,2017;Snell et al.,2017), addresses the few-shot problems, by extracting

Weak-shot Fine-grained Classiﬁcation via Similarity Transfer