Machine Learning
Machine learning is one of the core issues of artificial intelligence. By designing learning algorithms, computers have the ability to sense the external environment and model external systems. The main research directions of the institute include semi-supervised learning, indefinite kernel learning, deep learning, and sparse recovery.
Semi-supervised Learning
In data-driven machine learning methods, the labeling of data (such as the goal of regression problems, the category of classification problems, etc.) plays a very important role. However, in practical applications, the labeling of data is often difficult to obtain. For example, labels for medical imaging data are tagged by experienced physicians, subject to labor time and personal privacy, and often have few labeled samples for specific disease weights. Therefore, in the study of machine learning methods, it is necessary to consider the case of no label (unsupervised learning) or only a small number of labels (semi-supervised learning). With the support of the Natural Science Foundation of China, we continue to conduct research on semi-supervised learning. In the semi-supervised learning based on graphs, deep semi-supervised learning and other aspects have achieved certain innovation results and successfully applied in practical engineering.
Indefinite Kernel Learning
Kernel learning is an important machine learning method. Through the original-dual analysis, complex and unknown nonlinear mapping can be realized by designing a kernel function. Traditional kernel methods require that the kernel need to meet semi-definite conditions (in short, it needs to be semi-positive for generating the core matrix). In many applications, it is necessary to study the indefinite kernel learning method by breaking the semi-definite nature. For example, geodesics on manifolds often have good physical meaning, but distance-derived cores cannot satisfy semi-definite conditions. In addition, in the deep kernel learning method, it is necessary to train the function values of the kernel matrix, and this training makes the kernel unable to maintain positive definiteness. With the support of research funds such as the Thousand Talents Program, we continue to conduct research on the learning methods and training algorithms of the indefinite kernel, and obtain the indefinite kernel method, the multi-layer indeterminate kernel structure, and the indefinite kernel logistic regression under the framework of the least squares. Innovative results.
Sparse Signal/Image Learning
Sparse characteristics
are one of the essential attributes of big data. Research and discussion
on sparsity can help improve the accuracy of signal recovery, improve
the performance of medical image reconstruction such as CT, realize
image denoising super-resolution and network sparse coding. The
study of sparsity is also an important technique for controlling the
complexity of neural networks and improving the interpretability of
networks. In order to achieve sparsity, it is
necessary to study sparse machine learning methods, design non-convex
penalty functions to enhance sparsity, develop corresponding fast
optimization algorithms, and study the learning performance of sparse
algorithms. With the support of the Natural
Science Foundation and the Shanghai Science and Technology Commission,
we have conducted long-term research on image super-resolution,
non-convex sparse enhancement, one-bit compression sensing, and medical
image reconstruction. (RevisedTime:2018-11-12 11:23 Views:2482)
|
Copyright @2018 Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University.
Address: 2-427 SEIEE Building, 800 Dong Chuan Rd, Shanghai, 200240, P.R.China