Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss
[公告] 博士論文口試 陳以雷 (2014/5/1)
26 views
Skip to first unread message
二律背反
Apr 21, 2014, 2:59:17 AM4/21/14
to
清華大學資訊工程學系博士學位口試
學生姓名：陳以雷
指導教授：許秋婷教授
口試委員：
校內：賴尚宏教授、林嘉文教授、陳煥宗教授、許秋婷教授
校外：貝蘇章教授、廖弘源教授、莊仁輝教授、王聖智教授、簡仁宗教授
日期時間：103年5月1日(星期四) 10:00-12:00
口試地點：清華大學台達館601會議室
口試題目：Manifold Guided Tensor Completion under Low-rank Structure
論文摘要：
The success of research on matrix completion is evident in a variety of
real-world applications, such as computer vision, machine learning, and
knowledge mining. Tensor completion, which is a high-order extension of
matrix completion, has also generated a great deal of research interest in
recent years. Given a tensor with incomplete entries, existing methods
facilitate the ill-posed problem by assuming that the complete tensor
inherently exhibits low-rank structure. Predicting missing entries then boils
down to recovering a low-rank tensor from given entries. Factorization
schemes and completion schemes are two popular methodologies to characterize
the low tensor’s rank. In factorization schemes, an incomplete tensor is
drawn from a smaller core tensor under CP or Tucker model, while completion
schemes directly minimize the tensor’s rank using convex optimization.
However, as the number of missing entries increases, factorization schemes
tend to overfit the model structure due to their incorrectly predefined tensor
’s rank in practice, while completion schemes may fail to interpret the
model factors because they solely rely on rank minimization but disregard the
latent tensor structure. These two problems both limit the accuracy of tensor
recovery.
This dissertation introduces a novel concept to break the limitations in
current approaches: complete the missing entries and simultaneously capture
the underlying model structure. To this end, we propose a method called
Simultaneous Tensor Decomposition and Completion (STDC). Our major
contributions are two-fold. First, we leverage rank minimization techniques
with Tucker model decomposition in our STDC formulation; that is, we automate
rank estimation while carefully maintain the latent tensor structure. Second,
consider real-world tensor objects usually contain informative semantics of
their entries and in general the semantic relation encodes a joint-manifold
in low-dimensional space. We discover the latent manifold with a new
presented methodology, called Multilinear Graph Embedding (MGE), and study
its significance in tensor completion. As the model structure is implicitly
included in the Tucker model, we utilize MGE to encode factor priors, which
are usually known a priori as data semantics, to characterize the
joint-manifold drawn from the model factors. With the two contributions, our
method leverages two classic schemes, exploits the priors of low-rank
structure and low-dimensional manifolds, and thus accurately estimates the
model factors and missing entries.
Finally, because the proposed algorithm depends on a heuristic setting of
model parameters, we propose an effective strategy to reduce the number of
parameters so that only two hyper-parameters need to be determined. This
strategy assures the model parameters independent of the actual range of
given tensor, and also provides their physical meanings related to different
objectives. In addition, consider the factor priors are task-dependent and
could be unavailable. We propose a prior-free extension to automate the
joint-manifold learning and thereby enable STDC in dealing with unsupervised
applications. The extended approach is simply built on a smoothness prior of
general manifolds. With a set of augmented variables, we show that a new
methodology, called Permutation on Manifolds (PoM), is seamlessly included
into our STDC algorithm to approximate the smooth manifold’s shape.
We conducted experiments to empirically verify the convergence of our
algorithm on synthetic data, and evaluate its effectiveness on various kinds
of real-world data. The results demonstrate the efficacy of our method and
its potential usage in tensor-based applications. It also outperforms
state-of-the-art methods on multifactor data analysis and visual data
completion tasks.
主要著作：
Journal Paper
1. Y. L. Chen, C. T. Hsu, and H. Y. Mark Liao, “Simultaneous Tensor
Decomposition and Completion Using Factor Priors,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 36, no. 3, pp. 577-591, 2014.
2. Y. L. Chen and C. T. Hsu, “Multilinear Graph Embedding: Representation
and Regularization for Images,” IEEE Trans. Image Processing, vol. 23,
no. 2, pp. 741-754, 2014.
3. Y. L. Chen and C. T. Hsu, “Subspace Learning for Facial Age Estimation
Via Pairwise Age Ranking,” IEEE Trans. Information Forensic and Security,
vol. 8, no. 12, pp. 2164-2176, 2013.
4. Y. L. Chen and C. T. Hsu, “Detecting Recompression of JPEG Images via
Periodicity Analysis of Compression Artifacts for Tampering Detection,”
IEEE Trans. Information Forensic and Security, vol. 6, no. 2, pp. 396-406,
2011.
Conference Papers
1. Y. L. Chen and C. T. Hsu, “Implicit Rank-Sparsity Decomposition:
Applications to Saliency/Co-Saliency Detection,” ICPR, 2014. (accepted)
2. Y. L. Chen and C. T. Hsu, “A Generalized Low-Rank Appearance Model for
Spatio-Temporally Correlated Rain Streaks,” in Proc. ICCV, pp. 1968-1975,
2013.
3. Y. L. Chen and C. T. Hsu, “What Has Been Tampered? From A Sparse
Manipulation Perspective,” in Proc. MMSP, pp. 123-128, 2013.
4. Y. S. Lai, Y. L. Chen, and C. T. Hsu, “Single Image Dehazing With Optimal
Transmission Map,” in Proc.ICPR, pp. 388-391, 2012.
5. Y. L. Chen and C. T. Hsu, “Time-Variant Modeling for General Surface
Appearance,” in Proc. ICIP, pp. 1077-1080, 2011.
6. W. C. Chiou, Y. L. Chen, and C. T. Hsu, “Color Transfer for Complex
Content Images Based on Intrinsic Component,” in Proc. MMSP, pp. 156-161,
2010.
7. Y. L. Chen and C. T. Hsu, “Detecting Doubly Compressed Images Based on
Quantization Noise Model and Image Restoration,” in Proc. MMSP, pp. 1-6,
2009.
8. Y. L. Chen and C. T. Hsu, “Image Tampering Detection By Blocking
Periodicity Analysis in JPEG Compressed Images,” in Proc. MMSP, pp.
803-808, 2008.
--
[;32m※ Origin: 楓橋驛站<bbs.cs.nthu.edu.tw>
[;32m◆ From: fallcolor @ auth-gateway.cs.nthu.edu.tw [m
學生姓名：陳以雷
指導教授：許秋婷教授
口試委員：
校內：賴尚宏教授、林嘉文教授、陳煥宗教授、許秋婷教授
校外：貝蘇章教授、廖弘源教授、莊仁輝教授、王聖智教授、簡仁宗教授
日期時間：103年5月1日(星期四) 10:00-12:00
口試地點：清華大學台達館601會議室
口試題目：Manifold Guided Tensor Completion under Low-rank Structure
論文摘要：
The success of research on matrix completion is evident in a variety of
real-world applications, such as computer vision, machine learning, and
knowledge mining. Tensor completion, which is a high-order extension of
matrix completion, has also generated a great deal of research interest in
recent years. Given a tensor with incomplete entries, existing methods
facilitate the ill-posed problem by assuming that the complete tensor
inherently exhibits low-rank structure. Predicting missing entries then boils
down to recovering a low-rank tensor from given entries. Factorization
schemes and completion schemes are two popular methodologies to characterize
the low tensor’s rank. In factorization schemes, an incomplete tensor is
drawn from a smaller core tensor under CP or Tucker model, while completion
schemes directly minimize the tensor’s rank using convex optimization.
However, as the number of missing entries increases, factorization schemes
tend to overfit the model structure due to their incorrectly predefined tensor
’s rank in practice, while completion schemes may fail to interpret the
model factors because they solely rely on rank minimization but disregard the
latent tensor structure. These two problems both limit the accuracy of tensor
recovery.
This dissertation introduces a novel concept to break the limitations in
current approaches: complete the missing entries and simultaneously capture
the underlying model structure. To this end, we propose a method called
Simultaneous Tensor Decomposition and Completion (STDC). Our major
contributions are two-fold. First, we leverage rank minimization techniques
with Tucker model decomposition in our STDC formulation; that is, we automate
rank estimation while carefully maintain the latent tensor structure. Second,
consider real-world tensor objects usually contain informative semantics of
their entries and in general the semantic relation encodes a joint-manifold
in low-dimensional space. We discover the latent manifold with a new
presented methodology, called Multilinear Graph Embedding (MGE), and study
its significance in tensor completion. As the model structure is implicitly
included in the Tucker model, we utilize MGE to encode factor priors, which
are usually known a priori as data semantics, to characterize the
joint-manifold drawn from the model factors. With the two contributions, our
method leverages two classic schemes, exploits the priors of low-rank
structure and low-dimensional manifolds, and thus accurately estimates the
model factors and missing entries.
Finally, because the proposed algorithm depends on a heuristic setting of
model parameters, we propose an effective strategy to reduce the number of
parameters so that only two hyper-parameters need to be determined. This
strategy assures the model parameters independent of the actual range of
given tensor, and also provides their physical meanings related to different
objectives. In addition, consider the factor priors are task-dependent and
could be unavailable. We propose a prior-free extension to automate the
joint-manifold learning and thereby enable STDC in dealing with unsupervised
applications. The extended approach is simply built on a smoothness prior of
general manifolds. With a set of augmented variables, we show that a new
methodology, called Permutation on Manifolds (PoM), is seamlessly included
into our STDC algorithm to approximate the smooth manifold’s shape.
We conducted experiments to empirically verify the convergence of our
algorithm on synthetic data, and evaluate its effectiveness on various kinds
of real-world data. The results demonstrate the efficacy of our method and
its potential usage in tensor-based applications. It also outperforms
state-of-the-art methods on multifactor data analysis and visual data
completion tasks.
主要著作：
Journal Paper
1. Y. L. Chen, C. T. Hsu, and H. Y. Mark Liao, “Simultaneous Tensor
Decomposition and Completion Using Factor Priors,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 36, no. 3, pp. 577-591, 2014.
2. Y. L. Chen and C. T. Hsu, “Multilinear Graph Embedding: Representation
and Regularization for Images,” IEEE Trans. Image Processing, vol. 23,
no. 2, pp. 741-754, 2014.
3. Y. L. Chen and C. T. Hsu, “Subspace Learning for Facial Age Estimation
Via Pairwise Age Ranking,” IEEE Trans. Information Forensic and Security,
vol. 8, no. 12, pp. 2164-2176, 2013.
4. Y. L. Chen and C. T. Hsu, “Detecting Recompression of JPEG Images via
Periodicity Analysis of Compression Artifacts for Tampering Detection,”
IEEE Trans. Information Forensic and Security, vol. 6, no. 2, pp. 396-406,
2011.
Conference Papers
1. Y. L. Chen and C. T. Hsu, “Implicit Rank-Sparsity Decomposition:
Applications to Saliency/Co-Saliency Detection,” ICPR, 2014. (accepted)
2. Y. L. Chen and C. T. Hsu, “A Generalized Low-Rank Appearance Model for
Spatio-Temporally Correlated Rain Streaks,” in Proc. ICCV, pp. 1968-1975,
2013.
3. Y. L. Chen and C. T. Hsu, “What Has Been Tampered? From A Sparse
Manipulation Perspective,” in Proc. MMSP, pp. 123-128, 2013.
4. Y. S. Lai, Y. L. Chen, and C. T. Hsu, “Single Image Dehazing With Optimal
Transmission Map,” in Proc.ICPR, pp. 388-391, 2012.
5. Y. L. Chen and C. T. Hsu, “Time-Variant Modeling for General Surface
Appearance,” in Proc. ICIP, pp. 1077-1080, 2011.
6. W. C. Chiou, Y. L. Chen, and C. T. Hsu, “Color Transfer for Complex
Content Images Based on Intrinsic Component,” in Proc. MMSP, pp. 156-161,
2010.
7. Y. L. Chen and C. T. Hsu, “Detecting Doubly Compressed Images Based on
Quantization Noise Model and Image Restoration,” in Proc. MMSP, pp. 1-6,
2009.
8. Y. L. Chen and C. T. Hsu, “Image Tampering Detection By Blocking
Periodicity Analysis in JPEG Compressed Images,” in Proc. MMSP, pp.
803-808, 2008.
--
[;32m※ Origin: 楓橋驛站<bbs.cs.nthu.edu.tw>
[;32m◆ From: fallcolor @ auth-gateway.cs.nthu.edu.tw [m
0 new messages