Semi supervised, weakly-supervised, unsupervised, and active learning
An overview of semi supervised learning, weakly-supervised learning, unsupervised learning, and active learning.
Focused on recent deep learning-based image recognition approaches.
色々な定義
3
O. Chapelle, B.Schlkopf, and A. Zien, "Semi-Supervised Learning," in The MIT Press, 2010.
http://www.acad.bg/ebook/ml/MITPress-%20SemiSupervised%20Learning.pdf
4.
色々な定義
4
X. Zhu, "Semi-supervisedlearning literature survey," in Tech. rep. 1530, University of Wisconsin-Madison,
2005.
Pseudo label
Semi-supervised learning
• 予測対象そのものが教師として付加されたデータと、全く
教師が付加されていないデータで学習
•X. Zhu, "Semi-supervised learning literature survey,"
in Tech. rep. 1530, University of Wisconsin-Madison,
2005.
https://minds.wisconsin.edu/bitstream/handle/1793/60444/TR1530.pdf
• O. Chapelle, B. Schlkopf, and A. Zien, "Semi-Supervised
Learning," in The MIT Press, 2010.
http://www.acad.bg/ebook/ml/MITPress-%20SemiSupervised%20Learning.pdf
12
13.
Semi-supervised learning
• 考え方:近いやつは同じラベルやろ
•「近い」の定義が重要(manifoldをうまく捉えたり)
• Generative methods [Miller+,’96]
• Graph-based methods (label propagation) [Zhu+,’02]
• Low-density separation methods [Bennett+,’99]
13
D. Miller and H. Uyar, "A mixture of experts classifier with learning based on both labelled and unlabelled
data," in Proc. of NIPS, 1996.
X. Zhu and Z. Ghahramani, "Learning from labeled and unlabeled data with label propagation," in CMU
CALD tech. rep., 2002.
K. Bennett and A. Demiriz, "Semi-supervised support vector machines," in Proc. of NIPS, 1999.
MixMatch
• 異なるK個(=2, inpractice)のaugmentationを行ったラベルなし
データの推論結果をsharpenすることでpseudo labelを作成
• ラベルあり、なしデータ間のmixup
15
D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver, and C. Raffel, "MixMatch: A Holistic Approach
to Semi-Supervised Learning," in Proc. of NIPS, 2019.
https://github.com/google-research/mixmatch
ReMixMatch
• MixMatchをベースに改良
• Distributionalignment
• Augmentation anchoring
• CTAugment
• Rotation loss
(後述の弱教師)
17
D. Berthelot, N. Carlini, E. Cubuk, A. Kurakin, K. Sohn, H. Zhang, and C. Raffel, "ReMixMatch: Semi-
Supervised Learning with Distribution Matching and Augmentation Anchoring," in Proc. of ICLR, 2020.
https://github.com/google-research/remixmatch
C.f. AugMix
• オリジナルとaugmentした画像間の予測結果をconsistentにするロスを利用
25
D.Hendrycks, N. Mu, E. Cubuk, B. Zoph, J. Gilmer, and B. Lakshminarayanan, "AugMix: A Simple Data
Processing Method to Improve Robustness and Uncertainty," in Proc. of ICLR, 2020.
https://github.com/google-research/augmix
異なるOPを異なる回数かけた画像
の線形和と、更にオリジナルの線形
和でaugmentationを定義 通常のCE オリジナル、aug1, aug2間のJSD
26.
そもそもSSLの設定は現実的なのか?
• ラベルなし画像として、
ラベルあり画像のクラスを
過不足なく集めた画像集合を仮定
• どういう状況?!
•ラベルありとラベルなしの
クラスの分布が乖離すると性能低下
26
A. Oliver, A. Odena, C. Raffel, E. Cubuk, and I. Goodfellow, "Realistic Evaluation of Deep Semi-Supervised
Learning Algorithms," in Proc. of NIPS, 2018.
http://peluigi.hatenablog.com/entry/2018/09/07/162515(分かりやすい解説)
Weakly-supervised learning
• 予測対象の不完全な教師が付加されたデータから学習
•もともとはMultiple Instance Learning問題として提唱された
28
T. G. Dietterich, R. H. Lathrop, T. Lozano-Perez, "Solving the multiple instance problem with axis-parallel
rectangles," in Artificial Artificial Intelligence, Vol. 89, Issues 1-2, pp. 31-71, 1997.
• 通常は特徴量とラベルが1:1対応
• MILでは、複数の特徴量の集合(bag)に対して
ラベルが付く
• 分子形状が内部結合の状態によって変化するので
MILとして解いている
29.
Weakly-supervised learning
• 画像認識:1枚の画像を“bag” とみなし、画像のラベルに
対応する領域がどこかに1つ以上存在するという問題設定
29
O. Maron and A. Ratan, "Multiple-Instance Learning for Natural Scene Classification," in Proc. of ICML,
1998.
S. Andrews, I. Tsochantaridis, and T. Hofmann, "Support vector machines for multiple-instance learning," in
Proc. of NIPS, 2003.
30.
Weakly-supervised learning
• Multi-instancemulti-label learning (MIML) とかいうのも
30
Z. Zhou, M. Zhang, S. Huang, and Y. Li, "Multi-instance multi-label learning," in Artificial Intelligence, Vol.
176, Issue 1, pp. 2291-2320, 2012.
31.
Weakly-supervised learning
• 画像ラベルor BBOXからsegmentationを学習
31
G. Papandreou, L. Chen, K. Murphy, and A. Yuille, "Weakly- and Semi-Supervised Learning of a DCNN for
Semantic Image Segmentation," in Proc. of ICCV, 2015.
Unsupervised Learning
• 一般的には、クラスタリングや主成分分析等の
教師なしでデータ自体の構造や表現を獲得する手法
•DNNのコンテキストだと、教師なしで画像から
pretrainモデルを作る話(乱暴); self-supervisedとも
• 古典?:H. Barlow, "Unsupervised Learning," in Neural Computation, Vol.
1, Issue 3, pp. 295-311, 1989.
• DNN:Y. Bengio, A. Courville, and P. Vincent, "Representation Learning:
A Review and New Perspectives," in TPAMI, Vol. 35 Issue 8, pp. 1798-1828,
2013.
33
34.
Unsupervised Learning forDNN
• 自動的に教師となる情報を作り出せる
”pretext” taskで学習させる
• pretext tasks
• 画像の回転を推測させる
• パッチ間の位置関係を推測させる
• パッチのジグソーパズルを解かせる
• Contrastive loss系
• 基本的にsemi-supervisedとしても使える
34
D. Berthelot, N. Carlini, E. Cubuk, A. Kurakin, K. Sohn, H. Zhang, and C. Raffel, "ReMixMatch: Semi-
Supervised Learning with Distribution Matching and Augmentation Anchoring," in Proc. of ICLR, 2020.
35.
パッチ間の位置関係を推測させる
• 8-way classification
•問題を簡単に解けてしまう
“shortcut” の除去が重要
• エッジのつながり
→cropにjitter
• 画像の位置によって色ズレが発生する色収差
→グレーっぽくする
35
C. Doersch, A. Gupta, and A. Efros, "Unsupervised Visual Representation Learning by Context Prediction,"
in Proc. of ICCV, 2015.
画像の回転を推定させる
38
S. Gidaris, P.Singh, and N. Komodakis, "Unsupervised Representation Learning by Predicting Image
Rotations," in Proc. of ICLR, 2018.
39.
Contrastive loss系が流行り
• 同一画像を異なるaugmentation
したペアを同じ、他の画像ペアを違う
と当てられるかというpretexttask
39
(CPC) A. Oord, Y. Li, and O. Vinyals, "Representation Learning with Contrastive Predictive Coding," in
arXiv:1807.03748, 2018.
(CPCv2) O. Henaff, A. Srinivas, J. Fauw, A. Razavi, C. Doersch, S. Eslami, adn A. Oord, "Data-Efficient
Image Recognition with Contrastive Predictive Coding," in arXiv:1905.09272, 2019.
(MoCo) K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick, "Momentum Contrast for Unsupervised Visual
Representation Learning," in arXiv:1911.05722, 2019.
(PIRL) I. Misra and L. Maaten, "Self-Supervised Learning of Pretext-Invariant Representations," in
arXiv:1912.01991, 2019.
(SimCLR) T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, "A Simple Framework for Contrastive Learning
of Visual Representations," in arXiv:2002.05709, 2020.
Active learning
• ラベルなしデータにアノテーションをする際に、
このデータにラベルをつけるとモデルの精度が一番上がる
であろうデータを優先的にラベリングしていく考え方
41
M.Wang and X. Hua, "Active learning in multimedia annotation and retrieval: A survey," in ACM Trans. on
Intelligent Systems and TechnologyFebruary, 2011.
Y. Fu, X. Zhu, and B. Li , "A survey on instance selection for active learning," in Knowledge and Information
Systems, Vol. 35, pp 249–283, 2013.