24
Turning a Blind Eye: Explicit Removal of Biases and Variation from Deep Neural Network Embeddings Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics ECCV 2018 Workshop, 14/09/2018

Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Turning a Blind Eye:Explicit Removal of Biases and

Variation from Deep Neural Network Embeddings

Mohsan Alvi, Andrew Zisserman, and Christoffer NellåkerBias Estimation in Face Analytics

ECCV 2018 Workshop, 14/09/2018

Page 2: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Introduction

[email protected] 2

Convolutional Neural Networks are the state-of-the-art in image classification

Rely on “good” training data

Not clear what the network has

learned

Networks can cheat by learning

spurious variations

Lack of comprehensively annotated data

Page 3: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Contents

• Removing a bias from a network• Removing multiple spurious variations from a network• LAOFIW – Labeled Ancestral Faces in the Wild

[email protected] 3

Page 4: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Spurious Variations vs Biases

[email protected] 4

Spurious Variations

Biases

“Side information”

Page 5: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Gender Classification from Celebrity Faces

[email protected] 5

Page 6: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

IMDB Faces Dataset [1]

• Dataset consists of celebrity faces from International Movie DataBase• Contains Age, Gender, Identity Labels

• Created two subsets of the dataset• Age• Gender

• Cleaned labels

[email protected] 6

[1] Rothe, Timofte, Van Gool. (2015)

Page 7: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Biased Datasets

[email protected] 7

Page 8: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Extremely Biased Datasets

[email protected] 8

Training data

?

Test data

Page 9: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

ExtremelyBiased

Datasets

Only young women, only old men

[email protected] 9

Page 10: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Training on Extremely Biased DataEvaluated on a gender/age balanced test dataset

Gender classification accuracy: 70%

[email protected] 10

Page 11: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Turning a Blind Eye

• Primary task is the attribute of interest• Gender Classification

• Secondary task denotes bias to be unlearned• Age Classification

• Objective:learn feature representation that is informative for primary task, and uninformative for secondary task• Repurpose work in field of domain adaptation [2]

[email protected] 11

[2] Tzeng, Hoffman, Darrell, and Saenko. (2015)

Page 12: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Methods (1/3)

Based on VGG-M Network [2]

[2] Chatfield, Simonyan, Vedaldi, and Zisserman. (2014)

Minimize:

[email protected] 12

Page 13: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Methods (2/3)

Primary Loss

[email protected] 13

Minimize:

Page 14: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Methods (3/3)

Cross-entropy between classifier output and uniform distribution

[email protected] 14

Minimize:

Confusion Loss

Secondary Loss

Secondary Classification Secondary

Confusion

Act in opposition to each other

Alternate:

Primary Classification

&Secondary Confusion

Secondary Classification

Primary Classification

Secondary Confusion

Page 15: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Results – Removing a bias

Baseline Gender classification accuracy: 70%Age-Blind Gender classification accuracy: 86%

[email protected] 15

Page 16: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Removing multiple spurious variations (1/2)

Problem 1: Multiple biases may be present in dataset

[email protected] 16

AgeGender

Pose

Expression

Ancestry

Page 17: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Removing multiple spurious variations (2/2)

• Problem 2: no single dataset contains labels for all biases

• Each labeled for a single purpose• Leverage information from multiple datasets

[email protected] 17

Dataset 3:Ancestry

Dataset 1:Gender

Dataset 4:Pose

Dataset 2:Age

Page 18: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

14,000 images in four classes:• Sub-Saharan Africa• Western Europe• East Asia• Indian subcontinent

Labeled Ancestral Origin Faces in the

Wild

LAOFIW

[email protected] 18

Page 19: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Methods - Removing multiple spurious variations

For M secondary tasks:

[email protected] 19

Page 20: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Removing multiple spurious variations experiments

[email protected] 20

AncestryGender* PoseAge

GenderAncestry PoseAge

Experiment(1)

Experiment(2)

* Not extremely biased

Page 21: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Results - Removing multiple spurious variations (1/2)

Primary task: Gender Secondary tasks: Age, Ancestry, Pose

[email protected] 21

Page 22: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Results - Removing multiple spurious variations (2/2)

Primary task: Ancestry Secondary tasks: Age, Gender, Pose

[email protected] 22

Page 23: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Conclusions

• Can improve generalizability of models train on biased datasets• Can remove multiple spurious variations from feature representation

of network• LAOFIW – ancestral origin dataset

[email protected] 23

Page 24: Turning a Blind Eye - University of Oxfordvgg/publications/2018/Alvi18/... · 2018-10-04 · Mohsan Alvi, Andrew Zisserman, and Christoffer Nellåker Bias Estimation in Face Analytics

Questions?

[email protected] 24