An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning - - PowerPoint PPT Presentation

▶

Apr 02, 2023 37 likes •158 views

An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning Yaoyao Liu Bernt Schiele Qianru Sun MPI Informatics MPI Informatics Singapore Management University Research background Limitation : most algorithms are based on supervised

SLIDE 1

An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Yaoyao Liu MPI Informatics Bernt Schiele MPI Informatics Qianru Sun Singapore Management University

SLIDE 2

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Research background

Limitation: most algorithms are based on supervised learning,

so we need lots of labeled samples to train the model

(Image from Yao Lu)

Medical images: expensive to label the data Mitosis detection

SLIDE 3

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Few-shot learning: learning with limited data

Question: how to learn a model with limited labeled data? Task: few-shot image classification

Seen classes Many-shot Unseen classes Few-shot

(Images from Ravi and Larochelle)

SLIDE 4

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Review: meta-learning

Seen classes Unseen classes Training tasks Test task Meta-train Meta-test

(Images from Ravi, Larochelle, and Zhou)

SLIDE 5

Existing methods vs. our E3BM

Reference [1] Finn, Chelsea, Pieter Abbeel, and Sergey Levine. “Model-agnostic meta-learning for fast adaptation of deep networks.” ICML 2017; [2] Sun, Qianru, et al. “Meta-transfer learning for few-shot learning.” CVPR 2019; [3] Hu, Shell Xu, et al. “Empirical Bayes Transductive Meta-Learning with Synthetic Gradients.” ICLR 2020.

Existing methods:

A single base-learner
Arbitrary base-learning hyperparameters
Unstable

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Our E3BM:

An ensemble of multiple base-learners
Task-specific base-learning hyperparameters

+ Stable and robust (a) MAML [1]

θ

SGD SIB

θ

(c) SIB [3] (d) E3BM (ours)

θ

E3BM

(b) MTL [2]

SGD

SLIDE 6

meta update Learning rate Combination weight Base-learner initializer

... ...

Epoch-wise base-learner

Existing method: MAML[1]

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Reference [1] Finn, Chelsea, Pieter Abbeel, and Sergey Levine. “Model-agnostic meta-learning for fast adaptation of deep networks.” ICML 2017.

Predictions from a single base-learner Arbitrary base-learning hyperparameters For one training task:

SLIDE 7

deploy

...

meta update Learning rate Combination weight Base-learner initializer

... ... ...

deploy

...

Epoch-wise base-learner

Our method: E3BM framework

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Hyperprior Learner

...

Predictions from multiple base-learners Task-specific base-learning hyperparameters For one training task:

SLIDE 8

Two options of hyperprior learner

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

mean concat mean

FC FC

(a) Epoch-independent

For the m-th base epoch:

SLIDE 9

Two options of hyperprior learner

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

mean concat mean

LSTM

(b) Epoch-dependent

LSTM

For the m-th base epoch:

SLIDE 10

Boost the performance on THREE baselines

The 5-class few-shot classification results (%).

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

SLIDE 11

Values of generated and

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Settings: MTL+E3BM, ResNet-25, #base-learners = 100

SLIDE 12

Thank you!

An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning

Webpage: https://e3bm.yyliu.net/ Code: https://gitlab.mpi-klsb.mpg.de/yaoyaoliu/e3bm

Yaoyao Liu | An Ensemble of Epoch-wise Empirical Bayes for Few-Shot Learning