Lab Research Themes


   The main research interests of our lab are machine learning, especially deep learning and its applications. Our researches mainly focus on how to apply those technologies to prediction, classification, clustering and optimization in various real applications, such as pattern recognition and classification, system identification and control, data mining, financial analysis and so on. Based on the information extracted automatically from application domain, we are trying to develop application specific new efficient computational intelligent technologies, for examples, Quasi-linear support vector machines, Quasi-linear ARX neural networks, Deep neural networks, Deep kernel learning.

 

Methodology Researches:

  1. Quasi-Linear ARX Neural Networks and Its Applications to Control System
  2. Quasi-Linear Support Vector Machine and Its Applications to Nonlinear Classification
  3. Kernel Learning, Deep Kernel Learning and Applications

Application Researches:

  1. Feature Extraction, Feature Transform using AutoEncoder
  2. Object Detection, Key Point Detection, Pose Estimation
  3. Protein Function Predictions
  4. Text Classification, Relation Classification
  5. Applications of Deep learning + Reinforcement Learning
  6. Anomaly Detection

Quasi-Linear ARX Neural Networks and Its Applications to Control Systems


   The quasi-linear ARX Neural network is a class of neural network based nonlinear black-box models, which have not only a flexible representation ability, but also an easy-to-use structure. 1) It can be identified to be linear in the input variables. A nonlinear controller can be designed in the same way as a linear controller; 2) It can be identified to give a linear predictor and a nonlinear predictor at the same time. A switching control system can be designed to gurrantee the stability and to ensure the control accuracy. In this research, we study the modeling, the parameter estimation and the control system design.
  • I. Sutrisno, M.A. Jami'in, J. Hu and M.H. Marhaban, "A Self-Organizing Quasi-Linear ARX RBFN Model for Nonlinear Dynamical Systems Identification", SICE Journal of Control, Measurement, and System Integration, Vol.9, No.2, pp.70-77, 2016. doi
  • M.A. Jami'in, I. Sutrisno, J. Hu, N.B. Mariun and M.H. Marhaban, "Quasi-ARX Neural Network Based Adaptive Predictive Control for Nonlinear Systems", IEEJ Trans. on Electrical and Electronic Engineering, Vol.11, No.1, pp.83-90, 2016. doi
  • M.A. Jami'in, I. Sutrisno and J. Hu, "Maximum Power Tracking Control for a Wind Energy Conversion System Based on a Quasi-ARX Neural Network Model", IEEJ Trans. on Electrical and Electronic Engineering, Vol.10, No.4, pp.368-375, 2015. doi
  • I. Sutrisno, M.A. Jami'in and J. Hu, "An Improved Elman Neural Network Controller Based on Quasi-ARX Neural Network for Nonlinear Systems", IEEJ Trans. on Electrical and Electronic Engineering, Vol.9, No.5, pp.494-501, 2014. doi
  • L. Wang, Y. Cheng and J. Hu, "Stabilizing Switching Control for Nonlinear System Based on Quasi-ARX Model", IEEJ Trans. on Electrical and Electronic Engineering, Vol.7, No.4, pp.390-396, 2012. doi
  • Y. Cheng, L. Wang and J. Hu, "Identification of Quasi-ARX Neurofuzzy Model with an SVR and GA Approach", IEICE Trans. on Fundamentals of Electronics, communications and Computer Sciences, Vol.E95-A, No.5, pp.876-883, 2012. doi
  • Y. Cheng, L. Wang and J. Hu, "Identification of Quasi-ARX Neurofuzzy Model with an SVR and GA Approach", onlinear Theory and its Applications (NOLTA), IEICE, Vol.2, No.2, pp.165-179, 2011. doi
  • L. Wang, Y. Cheng and J. Hu, "A Quasi-ARX Neural Network with Switching Mechanism to Adaptive Control of Nonlinear Systems", SICE Journal of Control, Measurement, and System Integration, Vol.3, No.4, 2010. doi
  • 古月 敬之, "第二章 線形特性を有するニューラルネットワーク",  渡辺桂吾編著, 「ニューラルネットワーク計算知能」, pp.27-49, 森北出版株式会社, 東京, 2006
  • J.Hu and K.Hirasawa, "A Method for Applying Neural Networks to Control of Nonlinear Systems", in book entitled Neural Information Processing: Research and Development,  J.C.Rajapakse and L.Wang, Eds, pp..351-369, Springer, Berlin, GERMANY, 2004.
  • 古月・平澤, "非線形システムの制御のためのニューラルネットワーク予測モデル", 計測自動制御学会論文集, Vol.39, No.2, pp.168-175, 2003
  • J.Hu, K.Kumamaru and K.Hirasawa, " A Quasi-ARMAX Approach to the Modeling of Nonlinear Systems ", International Journal of Control, Vol.74, No.18, pp.1754-1766, 2001.
  • J.Hu, K.Kumamaru and K.Hirasawa, "A Neurofuzzy Approach to Fault Detection of Nonlinear Systems ",  Journal of Advanced Intelligence, Vol.3, No.6, pp.524-531,1999.
  • J.Hu, K.Hirasawa and K.Kumamaru, " A Neurofuzzy-Based Adaptive Predictor for Control of Nonlinear Systems ", Trans. of the Society of Instrument and Control Engineering, Vol.35, No.8, pp.1060-1068,1999.
  • J.Hu, K.Kumamaru, K.Inoue and K.Hirasawa, " KDI-Based Robust Fault Detection in Presence of Nonlinear Undermodeling", Trans. of the Society of Instrument and Control Engineering, Vol.35, No.2, pp.200-207,1999
  • J.Hu, K.Kumamaru, K.Inoue and K.Hirasawa, " A Hybrid Quasi-ARMAX Modeling Scheme for Identification of Nonlinear Systems ", Trans. of the Society of Instrument and Control Engineering, Vol.34, No.8, pp.977-985,1998.

Quasi-linear SVM and Its Application to Nonlinear Classification


  The quasi-linear support vector machine (SVM) is an SVM with a data-dependent quasi-linear kernel. A quasi-linear SVM classifier is constructed in two steps. In the first step, it builds a coarse model to approximate the nonlinear separation boundary by using multiple local linear models interpolated with a basis function such as radial basis function. In the second step, an SVM formulation is applied to implicitly optimizing the linear parameter set of coarse model. The SVM optimization is performed in an exact same way as a standard SVM with a quasi-linear kernel composed by using the interpolation functions.
  Quasi-Linear Kernel: It provides a scheme to compose a data-dependent kernel, which is less sensitive to noise data.
  Multi-Linear Feature Space: When using a RBF kernel, data points are mapped to a infinite high-dimensional feature space, while for the quasi-linear kernel, data points are mapped to a multi-linear feature space formed by multiple local linear models along the separation boundary.
  In this research, we study 1) how to detect local linear partitions along the separation boundary or to extract information of local linear partitions; 2) how to take advantages of the local linear partitions in solving problems of real world applications.
    a) A segmented approaches for imbalanced classification
    b) A coarse-to-fine approach for semi-supervised classification
    c) A segmented approaches for metric learning
  • P. Liang, W.Li and J.Hu, "Oversampling the Minority Class in a Multi-Linear Feature Space for Imbalanced Data Classification", IEEJ Trans. on Electrical and Electronic Engineering, Vol.13, No.10, 2018.doi
  • W.Li, B.Zhou, B.Chen and J.Hu, "A Geometry-Based Two-step Method for Nonlinear Classification Using Quasi-Linear Support Vector Machine", IEEJ Trans. on Electrical and Electronic Engineering, Vol.12, No.6, pp.883-890, 2017. doi
  • B.Zhou, W.Li and J.Hu, "A New Segmented Oversampling Method for Imbalanced Data Classification Using Quasi-Linear Support Vector Machine", IEEJ Trans. on Electrical and Electronic Engineering, Vol.12, No.6, pp.891-898, 2017. doi
  • B. Zhou, B. Chen and J. Hu, "Quasi-linear Support Vector Machine for Nonlinear Classification", IEICE Trans. on Fundamentals of Electronics, communications and Computer Sciences, Vol.E97-A, No.7, July, 2014. doi
  • B. Chen F. Sun and J. Hu, "Local Linear Multi-SVM Method for Gene Function Classification", in Proc. of the World Congress on Nature and Biologically Inspired Computing (NaBIC2010)(Kitakyushu), Dec., 2010, pp.183-188.

Object Detection, Key Point Detection, Pose Estimation


   Human keypoint estimation is a task to localize all visible key skeletons of humans in one static image. It is a fundamental research topic in the computer vision and a lot of researches are built upon it, like pose estimation and action recognition.
  A major characteristic of human keypoint estimation is dense prediction. Unlike a classification problem assigning a label to one image, it requires to assign a label to every pixel identifying whether it is a keypoint of humans and which skeleton it is.
  In practice, there are two major directions to make a good keypoint estimation. The first one is the bottom-up method. The algorithms categorized into this field first detect all possible human keypoints inside an image. Then, they find a way to combine all the detected keypoints to complete human skeletons in an instance level. The second one is the top-down method. A detection algorithm is first applied to detect all of humans inside an image. Then, keypoint estimation is applied on each human respectively. Both two directions have their own advantages. The bottom-up method holds its efficiency in speed, while the top-down method has a better performance.
  Currently, we are developing a new architecture of neural networks specific for the keypoint estimation. It has a large resolution of feature map even in deeper layers and keep a large receptive field at the same time. Besides, it maintains a comparable FLOPs (float operations) and is easy to be stacked as one single module.
  In previous research, the widely used neural networks is either not pre-trained by a large image dataset or not specifically designed for keypoint estimation. Therefore, our model fills this gap, that is directly designed for the keypoint estimation and easy to be pre-trained on large image classification problems like the ImageNet challenge. Moreover, we are developing and testing a series of mechanism to help the training process, such as online hard keypoint mining in the mini-batch level, and attention based heatmap regression.