Learning from conditional distributions via dual embeddings

Bo Dai, Niao He, Yunpeng Pan, Byron Boots, Le Song

Research output: Contribution to conferencePaper

Abstract

Many machine learning tasks, such as learning with invariance and policy evaluation in reinforcement learning, can be characterized as problems of learning from conditional distributions. In such problems, each sample x itself is associated with a conditional distribution p(z|x) represented by samples {zi}Mi=1, and the goal is to learn a function f that links these conditional distributions to target values y. These problems become very challenging when we only have limited samples or in the extreme case only one sample from each conditional distribution. Commonly used approaches either assume that z is independent of x, or require an overwhelmingly large set of samples from each conditional distribution. To address these challenges, we propose a novel approach which employs a new min-max reformulation of the learning from conditional distribution problem. With such new reformulation, we only need to deal with the joint distribution p(z, x). We also design an efficient learning algorithm, Embedding-SGD, and establish theoretical sample complexity for such problems. Finally, our numerical experiments, on both synthetic and real-world datasets, show that the proposed approach can significantly improve over existing algorithms.

Original languageEnglish (US)
StatePublished - Jan 1 2017
Event20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017 - Fort Lauderdale, United States
Duration: Apr 20 2017Apr 22 2017

Conference

Conference20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017
CountryUnited States
CityFort Lauderdale
Period4/20/174/22/17

Fingerprint

Reinforcement learning
Conditional Distribution
Invariance
Learning algorithms
Learning systems
Experiments
Reformulation
Reinforcement Learning
Min-max
Joint Distribution
Large Set
Learning
Learning Algorithm
Machine Learning
Extremes
Efficient Algorithms
Numerical Experiment
Target
Evaluation

ASJC Scopus subject areas

  • Artificial Intelligence
  • Statistics and Probability

Cite this

Dai, B., He, N., Pan, Y., Boots, B., & Song, L. (2017). Learning from conditional distributions via dual embeddings. Paper presented at 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, Fort Lauderdale, United States.

Learning from conditional distributions via dual embeddings. / Dai, Bo; He, Niao; Pan, Yunpeng; Boots, Byron; Song, Le.

2017. Paper presented at 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, Fort Lauderdale, United States.

Research output: Contribution to conferencePaper

Dai, B, He, N, Pan, Y, Boots, B & Song, L 2017, 'Learning from conditional distributions via dual embeddings' Paper presented at 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, Fort Lauderdale, United States, 4/20/17 - 4/22/17, .
Dai B, He N, Pan Y, Boots B, Song L. Learning from conditional distributions via dual embeddings. 2017. Paper presented at 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, Fort Lauderdale, United States.
Dai, Bo ; He, Niao ; Pan, Yunpeng ; Boots, Byron ; Song, Le. / Learning from conditional distributions via dual embeddings. Paper presented at 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, Fort Lauderdale, United States.
@conference{78aa20a8783d47f197e89d8b46a4c3ef,
title = "Learning from conditional distributions via dual embeddings",
abstract = "Many machine learning tasks, such as learning with invariance and policy evaluation in reinforcement learning, can be characterized as problems of learning from conditional distributions. In such problems, each sample x itself is associated with a conditional distribution p(z|x) represented by samples {zi}Mi=1, and the goal is to learn a function f that links these conditional distributions to target values y. These problems become very challenging when we only have limited samples or in the extreme case only one sample from each conditional distribution. Commonly used approaches either assume that z is independent of x, or require an overwhelmingly large set of samples from each conditional distribution. To address these challenges, we propose a novel approach which employs a new min-max reformulation of the learning from conditional distribution problem. With such new reformulation, we only need to deal with the joint distribution p(z, x). We also design an efficient learning algorithm, Embedding-SGD, and establish theoretical sample complexity for such problems. Finally, our numerical experiments, on both synthetic and real-world datasets, show that the proposed approach can significantly improve over existing algorithms.",
author = "Bo Dai and Niao He and Yunpeng Pan and Byron Boots and Le Song",
year = "2017",
month = "1",
day = "1",
language = "English (US)",
note = "20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017 ; Conference date: 20-04-2017 Through 22-04-2017",

}

TY - CONF

T1 - Learning from conditional distributions via dual embeddings

AU - Dai, Bo

AU - He, Niao

AU - Pan, Yunpeng

AU - Boots, Byron

AU - Song, Le

PY - 2017/1/1

Y1 - 2017/1/1

N2 - Many machine learning tasks, such as learning with invariance and policy evaluation in reinforcement learning, can be characterized as problems of learning from conditional distributions. In such problems, each sample x itself is associated with a conditional distribution p(z|x) represented by samples {zi}Mi=1, and the goal is to learn a function f that links these conditional distributions to target values y. These problems become very challenging when we only have limited samples or in the extreme case only one sample from each conditional distribution. Commonly used approaches either assume that z is independent of x, or require an overwhelmingly large set of samples from each conditional distribution. To address these challenges, we propose a novel approach which employs a new min-max reformulation of the learning from conditional distribution problem. With such new reformulation, we only need to deal with the joint distribution p(z, x). We also design an efficient learning algorithm, Embedding-SGD, and establish theoretical sample complexity for such problems. Finally, our numerical experiments, on both synthetic and real-world datasets, show that the proposed approach can significantly improve over existing algorithms.

AB - Many machine learning tasks, such as learning with invariance and policy evaluation in reinforcement learning, can be characterized as problems of learning from conditional distributions. In such problems, each sample x itself is associated with a conditional distribution p(z|x) represented by samples {zi}Mi=1, and the goal is to learn a function f that links these conditional distributions to target values y. These problems become very challenging when we only have limited samples or in the extreme case only one sample from each conditional distribution. Commonly used approaches either assume that z is independent of x, or require an overwhelmingly large set of samples from each conditional distribution. To address these challenges, we propose a novel approach which employs a new min-max reformulation of the learning from conditional distribution problem. With such new reformulation, we only need to deal with the joint distribution p(z, x). We also design an efficient learning algorithm, Embedding-SGD, and establish theoretical sample complexity for such problems. Finally, our numerical experiments, on both synthetic and real-world datasets, show that the proposed approach can significantly improve over existing algorithms.

UR - http://www.scopus.com/inward/record.url?scp=85032933472&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85032933472&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:85032933472

ER -