Representing and Explaining Novel Concepts with Minimal Supervision - - PowerPoint PPT Presentation

representing and explaining novel concepts with minimal
SMART_READER_LITE
LIVE PREVIEW

Representing and Explaining Novel Concepts with Minimal Supervision - - PowerPoint PPT Presentation

Representing and Explaining Novel Concepts with Minimal Supervision Dr. Zeynep Akata 2 April 2019 1 Outline Motivating the Importance of Side Information (Generalized) Zero-Shot Learning with Side Information Deeply Explainable Artificial


slide-1
SLIDE 1

Representing and Explaining Novel Concepts with Minimal Supervision

  • Dr. Zeynep Akata

2 April 2019

1

slide-2
SLIDE 2

Outline

Motivating the Importance of Side Information (Generalized) Zero-Shot Learning with Side Information Deeply Explainable Artificial Intelligence Summary and Future Work

2

slide-3
SLIDE 3

Outline

Motivating the Importance of Side Information (Generalized) Zero-Shot Learning with Side Information Deeply Explainable Artificial Intelligence Summary and Future Work

3

slide-4
SLIDE 4

Data Distribution in Large-Scale Datasets

Akata et.al. TPAMI’14

number of classes number of images

4

slide-5
SLIDE 5

Attributes as Side Information

Lampert et al. CVPR’09

class attributes images black-white has tail lives on land small gray has tail lives in water big zebra [1 0 1 1 0 1] whale [0 1 1 0 1 0]

5

slide-6
SLIDE 6

Attributes as Side Information

Lampert et al. CVPR’09

class attributes images black-white has tail lives on land small gray has tail lives in water big zebra [1 0 1 1 0 1] whale [0 1 1 0 1 0]

5

slide-7
SLIDE 7

Attributes as Side Information

Lampert et al. CVPR’09

class attributes images black-white has tail lives on land small gray has tail lives in water big zebra [1 0 1 1 0 1] whale [0 1 1 0 1 0]

5

slide-8
SLIDE 8

Zero-Shot Learning

images attributes ...

black-white has tail lives on land small

...

gray has tail lives in water big

...

black-white no tail lives on land medium white has tail lives on land tiny

6

slide-9
SLIDE 9

Muldimodal Embeddings

Akata et al. CVPR’13 & TPAMI’16

zebra whale

white black

IMAGES IMAGE FEATURES CLASS ATTRIBUTES CLASS LABELS

7

slide-10
SLIDE 10

Zero-Shot Learning Datasets

8

slide-11
SLIDE 11

Zero-Shot vs Generalized Zero-Shot Learning

Xian et al. CVPR 2017 Zero-Shot Learning Generalized Zero-Shot Learning CUB AWA CUB AWA Method u u u s H u s H Supervised Learning – – – 82.1 – – 96.2 –

slide-12
SLIDE 12

Zero-Shot vs Generalized Zero-Shot Learning

Xian et al. CVPR 2017 Zero-Shot Learning Generalized Zero-Shot Learning CUB AWA CUB AWA Method u u u s H u s H Supervised Learning – – – 82.1 – – 96.2 – Multimodal Embeddings 54.9 59.9 23.7 62.8 34.4 16.8 76.1 27.5

9

slide-13
SLIDE 13

Conclusions

Standard image classification models fail with the lack of labels

  • 1. Zero-Shot Learning is a challenging task that deserves attention
  • 2. Side information, e.g. attributes, is required to tackle zero-shot learning
  • 3. Several sources of side information exist: moving towards free-form text

Akata et.al. IEEE CVPR 2013, 2015, 2016 & IEEE TPAMI 2014, 2016

10

slide-14
SLIDE 14

Outline

Motivating the Importance of Side Information (Generalized) Zero-Shot Learning with Side Information Deeply Explainable Artificial Intelligence Summary and Future Work

11

slide-15
SLIDE 15

How to Tackle the Missing Data Problem?

Labels are difficult to obtain, attributes require expert knowledge

12

slide-16
SLIDE 16

How to Tackle the Missing Data Problem?

Labels are difficult to obtain, attributes require expert knowledge Proposed solution: Free text to image synthesis!

12

slide-17
SLIDE 17

Detailed Visual Descriptions

Reed et al. CVPR’16 The bird has a white underbelly, black feathers in the wings, a large wingspan, and a white beak. This bird has distinctive-looking brown and white stripes all over its body, and its brown tail sticks up. This swimming bird has a black crown with a large white strip on its head, and yellow eyes. This flower has a central white blossom surrounded by large pointed red petals which are veined and leaflike. Light purple petals with orange and black middle green leaves This flower is yellow and orange in color, with petals that are ruffled along the edges.

13

slide-18
SLIDE 18

Deep Representations of Text

Reed et al. CVPR’16

The beak is yellow and pointed and the wings are blue. Convolutional encoding Sequential encoding 14

slide-19
SLIDE 19

Text to Image Synthesis

This large bird has black feet and dark-brown feathers.

→ ??

15

slide-20
SLIDE 20

GAN1 Conditioned on Text

Reed et al. ICML’16 & NIPS’16

This flower has small, round violet petals with a dark purple center

φ φ

z ~ N(0,1)

This flower has small, round violet petals with a dark purple center

Generator Network Discriminator Network

φ(t) x := G(z,φ(t)) D(x’,φ(t))

1Generative Adversarial Networks [Goodfellow et al. NIPS’14]

16

slide-21
SLIDE 21

Text to Image Synthesis Results

a small sized bird that has tones of brown and dark red with a short stout bill the bird has a yellow bill, pink webbed feet, a white body with gray wings and gray tail feathers this bird is all blue, the top part of the bill is blue, but the bottom half is white. this is a large black bird with a pointy black beak this is a bird with a yellow belly, black head and breast and a black wing a vibrant colored bird of copper color, orange and blue with a very large orange bill Query Retrieval Query Retrieval Query Retrieval Query Generated Image Query Generated Image Query Generated Image

17

slide-22
SLIDE 22

Interpolatoing Between Sentences

‘Blue bird with black beak’ → ‘Red bird with black beak’ ‘Small blue bird with black wings’ → ‘Small yellow bird with black wings’ ‘This bird is bright.’ → ‘This bird is dark.’ ‘This bird is completely red with black wings’ ‘A small sized bird that has a cream belly and a short pointed bill’ ‘This is a yellow bird. The wings are bright blue’

18

slide-23
SLIDE 23

Generalized Zero-Shot Learning with Synthesized Images

CUB Data u s H Only real data 23.7 62.8 34.4

19

slide-24
SLIDE 24

Generalized Zero-Shot Learning with Synthesized Images

CUB Data u s H Only real data 23.7 62.8 34.4 With generated images 23.8 48.5 31.9 This is not better than having no images!

19

slide-25
SLIDE 25

f-CLSWGAN for Text to Image Feature Synthesis

Xian et al. CVPR’18

Head color: red Back color: black Crown color: red Wing shape: short

z ~ N(0, 1)

G(z, a) seen unseen

ResNet space

? x xg

f-CLSWGAN

CNN CNN

CNN feature space synthetic image real image

This is a small bird with a brown head and a yellow belly.

20

slide-26
SLIDE 26

Generalized Zero-Shot Learning with Synthesized Image Features

CUB Data u s H Only real data 23.7 62.8 34.4 With generated images 23.8 48.5 31.9

21

slide-27
SLIDE 27

Generalized Zero-Shot Learning with Synthesized Image Features

CUB Data u s H Only real data 23.7 62.8 34.4 With generated images 23.8 48.5 31.9 With generated features (f-CLSWGAN) 43.7 57.7 49.7

21

slide-28
SLIDE 28

CADA-VAE for Text to Image Feature Synthesis

Sch¨

  • nfeld et al. CVPR’19

E1 D1 E2 D2 red head pink belly brown wings gray beak E1 E2 D1 D2 red head pink belly brown wings gray beak 22

slide-29
SLIDE 29

CADA-VAE for Text to Image Feature Synthesis

Sch¨

  • nfeld et al. CVPR’19

E1 E2 D1 D2 E1 E2 D1 D2 E1 E2 D1 D2 E1 E2 D1 D2 E1 E2 D1 D2 E1 E2 D1 D2 E1 E2 D1 D2 COMPACT FIGURES (SMALL ENOUGH TO PUT 3 IN A ROW) SLIGHTLY MORE DETAILED FIGURES (PROBABLY TOO BIG TO PUT 3 IN A ROW) DETAILED FIGURE (THE EQUATIONS ON THE RIGHT ARE THE CROSS-RECONSTRUCTION LOSS. THE BASIC VAE LOSS IS NOT SHOWN)

CADA-VAE: DA-VAE: CA-VAE:

Current choice:

red head pink belly brown wings gray beak 22

slide-30
SLIDE 30

Generalized Zero-Shot Learning with Synthesized Image Features

CUB Data u s H Only real data 23.7 62.8 34.4 With generated images 23.8 48.5 31.9 With generated features (f-CLSWGAN) 43.7 57.7 49.7 With generated features (CADA-VAE) 63.6 51.6 52.4

23

slide-31
SLIDE 31

f-VAEGAN-D2 for Text to Image Feature Synthesis

Xian et al. CVPR’19

Encoder (E) Decoder/Generator(G)

Cape May Warbler Seen Feature Reconstruction (f-VAE)

Encoder (E) Decoder/Generator(G)

Cape May Warbler Discriminator1 (D1) Discriminator2 (D2)

VAE GAN D2

Transductive Learning (D2) Novel Feature Generation (f-WGAN) Seen Feature Reconstruction (f-VAE)

24

slide-32
SLIDE 32

f-VAEGAN-D2 for Text to Image Feature Synthesis

Xian et al. CVPR’19

Encoder (E) Decoder/Generator(G)

Cape May Warbler Discriminator1 (D1) Seen Feature Reconstruction (f-VAE) Novel Feature Generation (f-WGAN)

Encoder (E) Decoder/Generator(G)

Cape May Warbler Discriminator1 (D1) Discriminator2 (D2)

VAE GAN D2

Transductive Learning (D2) Novel Feature Generation (f-WGAN) Seen Feature Reconstruction (f-VAE)

24

slide-33
SLIDE 33

Generalized Zero-Shot Learning with Synthesized Image Features

CUB Data u s H Only real data 23.7 62.8 34.4 With generated images 23.8 48.5 31.9 With generated features (f-CLSWGAN) 43.7 57.7 49.7 With generated features (CADA-VAE) 63.6 51.6 52.4 With generated features (f-VAEGAN-D2) 63.2 75.6 68.9

25

slide-34
SLIDE 34

f-VAEGAN-D2 for Text to Image Feature Synthesis

Xian et al. CVPR’19

Encoder (E) Decoder/Generator(G)

Cape May Warbler Discriminator1 (D1) Seen Feature Reconstruction (f-VAE) Novel Feature Generation (f-WGAN)

Encoder (E) Decoder/Generator(G)

Cape May Warbler Discriminator1 (D1) Discriminator2 (D2)

VAE GAN D2

Transductive Learning (D2) Novel Feature Generation (f-WGAN) Seen Feature Reconstruction (f-VAE)

26

slide-35
SLIDE 35

f-VAEGAN-D2 for Text to Image Feature Synthesis

Xian et al. CVPR’19

Encoder (E) Decoder/Generator(G)

Cape May Warbler Discriminator1 (D1) Seen Feature Reconstruction (f-VAE) Discriminator2 (D2)

VAE GAN

Novel Feature Generation (f-WGAN)

D2

Transductive Learning (D2)

Encoder (E) Decoder/Generator(G)

Cape May Warbler Discriminator1 (D1) Discriminator2 (D2)

VAE GAN D2

Transductive Learning (D2) Novel Feature Generation (f-WGAN) Seen Feature Reconstruction (f-VAE)

26

slide-36
SLIDE 36

Generalized Zero-Shot Learning with Synthesized Image Features

CUB Data u s H Only real data 23.7 62.8 34.4 With generated images 23.8 48.5 31.9 With generated features (f-CLSWGAN) 43.7 57.7 49.7 With generated features (CADA-VAE) 63.6 51.6 52.4 With generated features (f-VAEGAN-D2) 63.2 75.6 68.9 With generated features (f-VAEGAN-D2 tran) 73.8 81.4 77.3

27

slide-37
SLIDE 37

Conclusions

Language complements visual information

  • 1. Provides an intuitive interface for the model
  • 2. Strong and generalizable: image classification & generation
  • 3. Many potential applications: Towards explaining what deep models think

Akata et al. IEEE CVPR 2013, 2015, 2016, TPAMI 2014, 2016 Reed et al. IEEE CVPR 2016 & ICML 2016 & NIPS 2016 Xian et al. IEEE CVPR 2016, 2017, 2018, 2019 Sch¨

  • nfeld et al. IEEE CVPR 2019; Dutta and Akata IEEE CVPR 2019

28

slide-38
SLIDE 38

Outline

Motivating the Importance of Side Information (Generalized) Zero-Shot Learning with Side Information Deeply Explainable Artificial Intelligence Summary and Future Work

29

slide-39
SLIDE 39

Human Machine Communication: Visual Question Answering

30

slide-40
SLIDE 40

Human Machine Communication: Visual Question Answering

What type of bird is this?

30

slide-41
SLIDE 41

Human Machine Communication: Visual Question Answering

What type of bird is this? It is a Cardinal What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings.

30

slide-42
SLIDE 42

Human Machine Communication: Visual Question Answering

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings.

30

slide-43
SLIDE 43

Generating Visual Explanations Model

Hendricks et al. ECCV’16

Explanation Sampler

This red bird has a red beak and a black face.

31

slide-44
SLIDE 44

Generating Visual Explanations Model

Hendricks et al. ECCV’16

Explanation Sampler

This red bird has a red beak and a black face. D: this bird has a white breast black wings and a red spot on its head. E: this is a white bird with a black wing and a black and white striped head. D: this bird has a white breast black wings and a red spot on its head. E: this is a black and white bird with a red spot on its crown. This is a Downy Woodpecker because... This is a Downy Woodpecker because...

31

slide-45
SLIDE 45

Generating Visual Explanations Model

Hendricks et al. ECCV’16

Explanation Sampler

This red bird has a red beak and a black face. D: this bird has a white breast black wings and a red spot on its head. E: this is a white bird with a black wing and a black and white striped head. D: this bird has a white breast black wings and a red spot on its head. E: this is a black and white bird with a red spot on its crown. This is a Downy Woodpecker because... This is a Downy Woodpecker because... Explanation: ...this is a brown and white spotted bird with a long pointed beak. Correct: Laysan Albatross, Predicted: Cactus Wren Correct & Predicted: Laysan Albatross Explanation: ...this bird has a white head and breast with a long hooked bill.

Cactus Wren Definition: ...this bird has a long thin beak with a brown body and black spotted feathers. Laysan Albatross Definition: ...this bird has a white head and breast a grey back and wing feathers and an orange beak.

31

slide-46
SLIDE 46

Human Machine Communication: Grounding & Counterfactuals

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings.

32

slide-47
SLIDE 47

Human Machine Communication: Grounding & Counterfactuals

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face

32

slide-48
SLIDE 48

Human Machine Communication: Grounding & Counterfactuals

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher?

32

slide-49
SLIDE 49

Human Machine Communication: Grounding & Counterfactuals

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings.

32

slide-50
SLIDE 50

Grounding Visual Explanations

Hendricks et al. ECCV’18

Explanation Sampler

This red bird has a red beak and a black face.

33

slide-51
SLIDE 51

Grounding Visual Explanations

Hendricks et al. ECCV’18

Explanation Sampler

attribute chunker This red bird has a red beak and a black face.

Explanation Grounder

red bird red beak black face

33

slide-52
SLIDE 52

Grounding Visual Explanations

Hendricks et al. ECCV’18

Explanation Sampler

attribute chunker attribute chunker This red bird has a red beak and a black face. This red bird has a black beak and a black face.

Explanation Grounder

red bird black beak black face red bird red beak black face

33

slide-53
SLIDE 53

Grounding Visual Explanations

Hendricks et al. ECCV’18

Explanation Sampler

1.02 attribute chunker 2.05 attribute chunker This red bird has a red beak and a black face. This red bird has a black beak and a black face.

Phrase-Critic Explanation Grounder

red beak black face red bird black face red bird black beak

red bird black beak black face red bird red beak black face

33

slide-54
SLIDE 54

Grounding Visual Explanations and Counterfactuals

This is a Red Winged Blackbird because …. this is a black bird with a red spot on its wingbars. Score: -11.29 this is a black bird with a red wing and a pointy black beak. This is a Red Faced Cormorant because …. this is a black bird with long neck and a red cheek patch. Score: -10.22 this is a black bird with a red cheek patch and a long white beak. This is a White Breasted Nuthatch because …. this is a white bird with a black crown and a black eye. Score: -13.20 this bird has a speckled belly and breast with a short pointy bill.

This bird is a Crested Auklet because this is a black bird with a small orange beak and it is not a Red Faced Cormorant because it does not have a long flat bill. This bird is a Parakeet Auklet because this is a black bird with a white belly and small feet and it is not a Horned Grebe because it does not have red eyes. This bird is a Least Auklet because this is a black and white spotted bird with a small beak and it is not a Belted Kingfisher because it does not have a long pointy bill.

34

slide-55
SLIDE 55

Grounding Visual Explanations and Counterfactuals

This is a Red Winged Blackbird because …. this is a black bird with a red spot on its wingbars. Score: -11.29 this is a black bird with a red wing and a pointy black beak. This is a Red Faced Cormorant because …. this is a black bird with long neck and a red cheek patch. Score: -10.22 this is a black bird with a red cheek patch and a long white beak. This is a White Breasted Nuthatch because …. this is a white bird with a black crown and a black eye. Score: -13.20 this bird has a speckled belly and breast with a short pointy bill.

Counterfactuals: Contrasting explanations are intuitive and informative This bird is a Crested Auklet because this is a black bird with a small orange beak and it is not a Red Faced Cormorant because it does not have a long flat bill. This bird is a Parakeet Auklet because this is a black bird with a white belly and small feet and it is not a Horned Grebe because it does not have red eyes. This bird is a Least Auklet because this is a black and white spotted bird with a small beak and it is not a Belted Kingfisher because it does not have a long pointy bill.

34

slide-56
SLIDE 56

Textual Explanations for Self Driving Vehicles

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings. What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why do you think so? Because there are no cars in its lane. What is happening here? The car is moving forward.

35

slide-57
SLIDE 57

Textual Explanations for Self Driving Vehicles

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings. What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why do you think so? Because there are no cars in its lane. What is happening here? The car is moving forward.

35

slide-58
SLIDE 58

Textual Explanations for Self Driving Vehicles

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings. What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why do you think so? Because there are no cars in its lane. What is happening here? The car is moving forward.

35

slide-59
SLIDE 59

Textual Explanations for Self Driving Vehicles

What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why not a Vermilion Flycatcher? It is not a Vermilion Flycatcher because it does not have black wings. What type of bird is this? It is a Cardinal because it is a red bird with a red beak and a black face Why do you think so? Because there are no cars in its lane. What is happening here? The car is moving forward.

35

slide-60
SLIDE 60

Driving Explanations Model

Kim et al. ECCV’18

spatial attention Vehicle Controller The car is moving forward because there are no cars in its lane. spatial attention Vehicle Controller Explanation with Strongly Aligned Attention (SAA) spatio-temporal attention Explanation with Weakly Aligned Attention (WAA) The car is moving forward because there is no stop sign. temporal attention attention alignment loss

36

slide-61
SLIDE 61

Driving Explanations Model

Kim et al. ECCV’18

spatial attention The car is moving forward because there are no cars in its lane. Vehicle Controller Explanation with Strongly Aligned Attention (SAA) temporal attention The car is moving forward because there are no cars in its lane. spatial attention Vehicle Controller Explanation with Strongly Aligned Attention (SAA) spatio-temporal attention Explanation with Weakly Aligned Attention (WAA) The car is moving forward because there is no stop sign. temporal attention attention alignment loss

36

slide-62
SLIDE 62

Driving Explanations Model

Kim et al. ECCV’18

spatial attention The car is moving forward because there are no cars in its lane. Vehicle Controller Explanation with Strongly Aligned Attention (SAA) spatio-temporal attention Explanation with Weakly Aligned Attention (WAA) The car is moving forward because there is no stop sign. temporal attention attention alignment loss The car is moving forward because there are no cars in its lane. spatial attention Vehicle Controller Explanation with Strongly Aligned Attention (SAA) spatio-temporal attention Explanation with Weakly Aligned Attention (WAA) The car is moving forward because there is no stop sign. temporal attention attention alignment loss

36

slide-63
SLIDE 63

Driving Explanations Results

The car heads down the road because traffic is moving at a steady pace. The car is slowing because it is approaching a stop sign. The car is stopped because the car in front of it is stopped.

37

slide-64
SLIDE 64

Zero-Shot Explanations

Xian et al. CVPR’19

… it has a brown center and yellow petals .

Textual Explanation

… because it has a brown center and yellow petals . This flower has a large brown center and its petals are long.

f-VAEGAN-D2

Sunflower Visualization

Transductive Learning (D2)

Feature Generator (f-WGAN) Feature Reconstruction (f-VAE)

real image ∈

This is a sunflower because ... feature space … this flower has a wide brown center and tapered yellow petals. … this flower has a wide center and layers of wide, tapered yellow petals. This is a Sunflower because ... … this flower has petals that are white and has a bushy yellow center … the flower is big with white petals, and a bulb of yellow colored anthers. This is a Tree Poppy because ... … this flower has simple rows of

  • verlapping orange petals with a

notched tip of yellow stamen in the center.

Seen Classes Unseen Classes

This is a Marigold because ... … this flower has layers of long tapered pale yellow petals surrounding orange and red stamen. … this flower is pink in color, and has petals that are drooping downward. … this flower has pink petals that are pointed down, and a lot of red stamen in the center This is a Purple Coneflower because ...

R S R S

… this flower has red petals that have yellow tips. … this flower has petals that are red with yellow edges This is a Blanket Flower because ... … the petals of the flower are light pink, while the anthers are white and yellow. … this flower is pink and white in color, with petals that are rounded. This is a Pink Primrose because ... … the petals on this flower are mostly lavender in color and the inner stamen is the color purple. … this flower is green, white, and purple in color, and has petals that are oval shaped. This is a Passion Flower because ... … this flower has petals that are red with pointy tips … this flower has a lot of very thin red petals and a lot of white stamen on it This is a King Protea because ...

Challenging Classes

R S

… this flower has wide trumpet shaped purple flowers with a star shape. This is a Canterburry Bells because … … this flower has broad alternating leaves, and its pink colored petals are lighter pink. This is a Sweat Pea because … … the flowers color of the flower are visible. The stamen and pistil <unk> from it. This is a Balloon Flower because … … this flower has petals that are pink and white with green pedicel. … the petals on this flower are mostly bulb shaped purple. This is a Cameilla because … … the flower has five purple petals with white stamen and a white pistil. … this red flower has rounded petals and yellow stamen with yellow anthers. … the petals of the flower are layered in layers while the anthers and are yellow in color.

38

slide-65
SLIDE 65

Zero-Shot Explanations

Xian et al. CVPR’19

… it has a brown center and yellow petals .

Textual Explanation

… because it has a brown center and yellow petals . This flower has a large brown center and its petals are long.

f-VAEGAN-D2

Sunflower Visualization

Transductive Learning (D2)

Feature Generator (f-WGAN) Feature Reconstruction (f-VAE)

real image ∈

This is a sunflower because ... feature space … this flower has a wide brown center and tapered yellow petals. … this flower has a wide center and layers of wide, tapered yellow petals. This is a Sunflower because ... … this flower has petals that are white and has a bushy yellow center … the flower is big with white petals, and a bulb of yellow colored anthers. This is a Tree Poppy because ... … this flower has simple rows of

  • verlapping orange petals with a

notched tip of yellow stamen in the center.

Seen Classes Unseen Classes

This is a Marigold because ... … this flower has layers of long tapered pale yellow petals surrounding orange and red stamen. … this flower is pink in color, and has petals that are drooping downward. … this flower has pink petals that are pointed down, and a lot of red stamen in the center This is a Purple Coneflower because ...

R S R S

… this flower has red petals that have yellow tips. … this flower has petals that are red with yellow edges This is a Blanket Flower because ... … the petals of the flower are light pink, while the anthers are white and yellow. … this flower is pink and white in color, with petals that are rounded. This is a Pink Primrose because ... … the petals on this flower are mostly lavender in color and the inner stamen is the color purple. … this flower is green, white, and purple in color, and has petals that are oval shaped. This is a Passion Flower because ... … this flower has petals that are red with pointy tips … this flower has a lot of very thin red petals and a lot of white stamen on it This is a King Protea because ...

Challenging Classes

R S

… this flower has wide trumpet shaped purple flowers with a star shape. This is a Canterburry Bells because … … this flower has broad alternating leaves, and its pink colored petals are lighter pink. This is a Sweat Pea because … … the flowers color of the flower are visible. The stamen and pistil <unk> from it. This is a Balloon Flower because … … this flower has petals that are pink and white with green pedicel. … the petals on this flower are mostly bulb shaped purple. This is a Cameilla because … … the flower has five purple petals with white stamen and a white pistil. … this red flower has rounded petals and yellow stamen with yellow anthers. … the petals of the flower are layered in layers while the anthers and are yellow in color.

38

slide-66
SLIDE 66

Conclusions

Generating visual/textual explanations

  • 1. A means for model interpretation: necessary to improve deep models
  • 2. Important criteria to trust deep models: through explanations
  • 3. A step towards effective human-machine communication

Hendricks et al. ECCV 2016 & ECCV 2018, Park et al. IEEE CVPR 2018, Kim et al. ECCV 2018 Xian et al. IEEE CVPR 2019

39

slide-67
SLIDE 67

Outline

Motivating the Importance of Side Information (Generalized) Zero-Shot Learning with Side Information Deeply Explainable Artificial Intelligence Summary and Future Work

40

slide-68
SLIDE 68

Summary

  • 1. Multi-modal Joint Embeddings tackle lack of visual data

[Akata et al. TPAMI’14 & CVPR’13, CVPR’15 & TPAMI’16, Xian et al. CVPR’16]

41

slide-69
SLIDE 69

Summary

  • 1. Multi-modal Joint Embeddings tackle lack of visual data

[Akata et al. TPAMI’14 & CVPR’13, CVPR’15 & TPAMI’16, Xian et al. CVPR’16]

  • 2. Vision and Language complement each other for generating novel concepts

[Reed et al. CVPR’16 & ICML’16 & NIPS’16, Xian et al. CVPR’18 & CVPR’19, Sch¨

  • nfeld et al. CVPR’19, Dutta and Akata CVPR’19 ]

41

slide-70
SLIDE 70

Summary

  • 1. Multi-modal Joint Embeddings tackle lack of visual data

[Akata et al. TPAMI’14 & CVPR’13, CVPR’15 & TPAMI’16, Xian et al. CVPR’16]

  • 2. Vision and Language complement each other for generating novel concepts

[Reed et al. CVPR’16 & ICML’16 & NIPS’16, Xian et al. CVPR’18 & CVPR’19, Sch¨

  • nfeld et al. CVPR’19, Dutta and Akata CVPR’19 ]
  • 3. Developing explainable deep models is important for user acceptance

[Hendricks et al. ECCV’16 & ECCV’18, Park et al. CVPR’18, Kim et al. ECCV’18, Xian et.al. CVPR’19]

41

slide-71
SLIDE 71

Future of Deeply Explainable Artificial Intelligence The car is driving down an empty road. A ball is on the side of the road. The car should slow down. The ball is in the middle of the road. There is a child running towards the ball. The car should stop and not hit the child.

42

slide-72
SLIDE 72

Future of Deeply Explainable Artificial Intelligence The car is driving down an empty road. A ball is on the side of the road. The car should slow down. The ball is in the middle of the road. There is a child running towards the ball. The car should stop and not hit the child.

User: What happened?

42

slide-73
SLIDE 73

Future of Deeply Explainable Artificial Intelligence The car is driving down an empty road. A ball is on the side of the road. The car should slow down. The ball is in the middle of the road. There is a child running towards the ball. The car should stop and not hit the child.

User: What happened? XAI: I was driving down an empty road. I decided to slow down as a ball appeared on the right. I saw a child running towards the ball, so I decided to stop.

42

slide-74
SLIDE 74

Future of Deeply Explainable Artificial Intelligence The car is driving down an empty road. A ball is on the side of the road. The car should slow down. The ball is in the middle of the road. There is a child running towards the ball. The car should stop and not hit the child.

User: What happened? XAI: I was driving down an empty road. I decided to slow down as a ball appeared on the right. I saw a child running towards the ball, so I decided to stop. User: What would have happened if you did not stop ?

42

slide-75
SLIDE 75

Future of Deeply Explainable Artificial Intelligence The car is driving down an empty road. A ball is on the side of the road. The car should slow down. The ball is in the middle of the road. There is a child running towards the ball. The car should stop and not hit the child.

User: What happened? XAI: I was driving down an empty road. I decided to slow down as a ball appeared on the right. I saw a child running towards the ball, so I decided to stop. User: What would have happened if you did not stop ? XAI: If there was an impact, the child would have gotten hurt.

42

slide-76
SLIDE 76

Thank you!

43