Impact of ImageNet Model Selection on Domain Adaptation
Youshan Zhang & Brian D. Davison
03/01/2020
Computer Science & Engineering
Outline
5
2
3
4
Problem
Methods
State-of-the-art methods
1 Introduction
Conclusion & Future work
Datasets & Results
6
1
Introduction
Background
Ø Big data & different types of data
Ø Original elementary form & not labeled
Ø Manually label data: time-consuming & expensive
Ø Model reuse based on labeled data
§ Conventional machine learning: lower accuracy
1
2Text Image Audio
What is domain adaptation (DA)?
Human: previous experience and knowledge for reasoning & learning
Machine: apply knowledge from other fields into current applications
Key idea:
Ø How to find similar domain knowledge for transferring?
1
3
Why DA ?
Label data: time-consuming & expensive
Train from scratch: tedious
Design customized model: complex
Data shift/bias
1
4Source Domain Target Domain
Outline
5
2
3
4
Problem
Methods
State-of-the-art methods
1 Introduction
Conclusion & Future work
Datasets & Results
6
5
Problem2
6
XT = { (xj, ?) | j = 1, 2, … , nt}
XS = { (xi, yi) | i = 1, 2, … , ns}
Outline
5
2
3
4
Problem
Methods
State-of-the-art methods
1 Introduction
Conclusion & Future work
Datasets & Results
6
7
State-of-the-art methods3
8
Traditional methods
Ø Feature selection[Blitzer et al., 2006; Long et al., 2014]
Ø Subspace learning[Gopalan et al., 2011; Gonget al., 2012; Zhang et al., 2019b]
Ø Distribution adaptation[Panet al., 2011; Jiang et al., 2017; Wang et al., 2018]
Deep learning based methods
Ø Discrepancy based[Tzeng et al., 2014; Long et al.,2015; Ghifary et al., 2015]
Ø Reconstruction based[Bousmalis et al., 2016]
Ø Adversarial learning based[Ganinet al., 2016; Tzeng et al., 2017; Liu et al., 2019]
Limitations3
9
More or less rely on the backbone networks
Not explore other ImageNet models
Not know which is the best layer
Outline
5
2
3
4
Problem
Methods
State-of-the-art methods
1 Introduction
Conclusion & Future work
Datasets & Results
6
10
ImageNet Models4
11
ImageNet Models4
12
Extracted features visualization 4
13
Domain adaptation methods4
Support vector machines (SVM) & 1-nearest neighbor (1NN)
Geodesic Flow Kernel (GFK) & Geodesic sampling on manifolds (GSM)
CORrelation Alignment (CORAL)
Transfer Joint Matching (TJM)
Balanced distribution adaptation(BDA) & Joint distribution alignment (JDA) & Joint Geometrical and Statistical Alignment (JGSA) & Adaptation Regularization (ARTL) & Manifold Embedded Distribution Alignment (MEDA) & Modified Distribution Alignment (MDA)
14
Significance analysis 4
Correlation coefficient
Coefficient of determination
Ø The higher the better
15
Outline
5
2
3
4
Problem
Methods
State-of-the-art methods
1 Introduction
Conclusion & Future work
Datasets & Results
6
16
Datasets & Results5
17
Datasets
Results5
18
Office + Caltech 10
Results5
19
Office-Home
Results5
20
Office31
Classification Accuracy5
21
Classification Accuracy5
22
Classification Accuracy5
23
Best feature extraction layer5
24
Take home messages5
Features from a higher-performing ImageNet-trained model are more valuable than those from a lower-performing model for unsupervised domain adaptation
The layer prior to the last fully connected layer is the best layer for feature extraction
25
Conclusion & future work6
We are the first to examine how features from many different ImageNet models affect domain adaptation
Search the best architecture for feature extraction
Feature fusion
26
6
27
Thank you!
Questions?