Skip to main content

Table 4 Results on the standard ImageNet ZSL test splits

From: Semantic embeddings of generic objects for zero-shot learning

   

2-hop

3-hop

All

ZSL module

Description

Semantic module

Top-1

Top-5

Top-10

Top-1

Top-5

Top-10

Top-1

Top-5

Top-10

Ridge reg.

WordNet lemmas

word2vec

7.66

21.00

29.90

3.08

9.35

13.78

0.89

2.70

4.23

  

FastText

12.98

32.35

41.68

2.96

9.01

13.20

1.30

4.00

6.01

  

Glove

13.47

32.96

42.99

3.08

9.35

13.78

1.34

4.15

6.30

Ridge reg.

WordNet graph

TransE

5.77

8.73

10.16

1.07

1.71

1.99

0.42

0.65

0.76

  

DistMult

16.94

37.61

43.85

3.28

9.57

12.39

1.35

3.91

5.08

  

TransE∗

20.13

48.32

58.06

3.65

11.84

17.05

1.51

4.90

7.21

  

ConvE

3.23

9.12

12.46

1.30

2.14

3.26

0.42

1.72

3.10

  

Poincarre

11.81

28.86

37.53

2.02

5.93

8.79

0.79

2.32

3.46

Ridge reg.

BabelNet graph

TransE

2.82

5.11

7.16

1.03

1.41

1.75

0.37

0.88

1.01

  

DistMult

8.42

20.31

27.33

1.82

5.14

7.64

0.78

2.23

3.40

  

TransE∗

17.76

42.46

53.47

3.62

10.82

15.65

1.53

4.69

6.97

Ridge reg.

WordNet definitions

InferSent

4.06

12.37

18.52

1.18

3.91

6.15

0.49

1.66

2.67

  

DictRep

6.06

18.74

27.28

1.52

5.58

9.05

0.63

2.32

3.84

  

CapRep

3.45

10.86

16.37

1.13

2.97

4.35

0.21

0.56

1.01

  

Sent2vec

5.93

17.57

25.57

1.65

5.60

8.92

0.67

2.36

3.85

  

FastSent

1.82

5.31

9.86

0.82

2.11

3.21

0.19

0.43

0.75

  

SkipThought

0.50

1.38

2.11

0.17

0.46

0.67

0.06

0.17

0.26

Ridge reg.

Wikipedia articles

TFIDF

9.03

26.53

37.31

-

-

-

-

-

-

State of the art

SYNC [6]

WordNet lemmas

word2vec

9.26

-

-

2.29

-

-

0.96

-

-

CONSE [7]

  

7.63

  

2.18

  

0.95

  

ESZSL [36]

  

6.35

  

1.51

  

0.62

  

ALE [16]

  

5.38

  

1.32

  

0.5

  

LATEM [15]

  

5.45

  

1.32

  

0.5

  

SJE [14]

  

5.31

  

1.33

  

0.52

  

DEVISE [5]

  

5.25

  

1.29

  

0.49

  

CMT [37]

  

2.88

  

0.67

  

0.29

  

GCNZ [9]

Lemmas and graph

Glove&GCN

19.8

53.2

65.4

4.1

14.2

20.2

1.8

6.3

9.1

ADGPM [8]

  

26.6

60.3

72.3

6.3

19.3

27.7

3.0

9.3

13.9

  1. The upper part of the table shows our results using a ridge regression model with different semantic representations. The bottom part of the table shows the state-of-the-art results as reported in [4], with the additional entries of [8, 9]. Bold entries represent the best results obtained in each category