ABCDEFGHIJKLMNOPQRSTUVWXYZ
1
Multitask learning
2
2017-05-21Learning to Multi-Task by Active Sampling
https://arxiv.org/pdf/1702.06053.pdf
Multi-tasking on 21 Atari games simultaneously. They outperform humans on 8 games: (1) Crazy Climber (2) Krull (3) Demon Attack (4) Gopher (5) Breakout (6) Assault (7) Star Gunner (8) Tutankham |||||| They reach >30% of human performance on 6 games: (1) Name This Game (88%) (2) Bank Heist (85%) (3) Phoenix (60%) (4) Space Invaders (34%) (5) Amidar (32%) (6) Centipede (30%) |||||| They fail to reach >30% of human performance on 7games: (7 games): (1) Chopper Command (29%) (2) Alien (23%) (3) Asterix (20%) (4) Bowling (10%) (5) Beam Rider (8%) (6) Seaquest (5%) (7) Kangaroo (0.7%)
3
2017-06-16One Model To Learn Them All
https://arxiv.org/pdf/1706.05137.pdf
Simultaneously on: (1) ImageNet (2) speech recognition (WSJ) (3) WMT En<=>Ge, En<=>Fr translation (4) COCO image captioning
4
2016-09-07
UberNet : Training a ‘Universal’ Convolutional Neural Network for Low-, Mid-, and High-Level Vision using Diverse Datasets and Limited Memory
https://arxiv.org/pdf/1609.02132.pdf
"In this work we introduce <...> a ‘swiss knife’ for vision tasks; we call this architecture an UberNet <...> We obtain competitive performance while jointly addressing all of these tasks in 0.7 seconds per frame on a single GPU. A demonstration of this system can be found at cvn.ecp.fr/ubernet/ "
5
2017-05-22Learning multiple visual domains with residual adapters
https://arxiv.org/pdf/1705.08045.pdf
When their NN is trained, you give only image from one of 10 datasets and it automatically classifies what dataset it's from, adjusts 1% of its parameters (99% params are shared for all datasets) and gives not so less performance compared to SOTA on the dataset. While training, however, you give to NN both image and # of dataset it's taken from. They name their collection of 10 datasets as a visual decathlon challenge
6
2017-05-11Incremental Learning Through Deep Adaptation
https://arxiv.org/pdf/1705.04228.pdf
8 image datasets adding only 22% of params close to sotas CIFAR-10, Omniglot, SVHN...
7
2017-07-21Learning Transferable Architectures for Scalable Image Recognition
https://arxiv.org/pdf/1707.07012.pdf
8
2017-05-05
Supervised Learning of Universal Sentence Representations from Natural Language Inference Data
https://arxiv.org/pdf/1705.02364.pdf
we show how universal sentence representations trained using the supervised data of the Stanford Natural Language Inference dataset can consistently outperform unsupervised methods like SkipThought
9
2017-06-03See, Hear, and Read: Deep Aligned Representations
https://arxiv.org/pdf/1706.00932.pdf
mirror neurons: http://people.csail.mit.edu/yusuf/see-hear-read/
10
2017-10-27
Generalization Tower Network: A Novel Deep Neural Network Architecture for Multi-Task Learning
https://arxiv.org/pdf/1710.10036.pdf
simultaneously on 18 shooting games of Atari ~50% of single-task
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100