摘要 :
The classical algorithm ISOMAP can find the intrinsic low-dimensional structures hidden in high-dimensional data uniformly distributed on or around a single manifold. But if the data are sampled from multi- class, each of which co...
展开
The classical algorithm ISOMAP can find the intrinsic low-dimensional structures hidden in high-dimensional data uniformly distributed on or around a single manifold. But if the data are sampled from multi- class, each of which corresponds to an independent manifold, and clusters formed by data points belonging to each class are separated away, several disconnected neighborhood graphs will occur, which leads to the failure of ISOMAP. Moreover, ISOMAP behaves in an unsupervised manner and therefore works less effectively for classification. In this paper, two improved versions of ISOMAP, namely Multi-Class Multi-Manifold ISOMAP (MCMM-ISOMAP) for data visualization and ISOMAP for Classification (ISOMAP-C), are proposed respectively. MCMM-ISOMAP constructs a single neighborhood graph, named a between-class neighborhood graph by connection of between-class points with shortest distance of each within-class neighborhood graph, and then ISOMAP algorithm is applied to find the intrinsic low-dimensional embedding structure. ISOMAP-C is essentially an extension of MCMM-ISOMAP to a supervised manner, which is multiplied by scaling factor greater than one so that low dimensional data set after mapping become more compact within class and more separate between classes. Finally, the mapping function from original high dimensional space to low dimensional space can be approximately modeled using Back-Propagation neural network combined with genetic algorithm. Experimental results using MCMM-ISOMAP on synthetic and real data reveal its effectiveness and ones using ISOMAP-C show that the performance is greatly enhanced and robust to noisy data.
收起
摘要 :
Segmentation of skin lesion is an important step in the overall automated diagnostic systems used for early detection of skin cancer. Skin lesions can have various different forms which makes segmentation a difficult and complex t...
展开
Segmentation of skin lesion is an important step in the overall automated diagnostic systems used for early detection of skin cancer. Skin lesions can have various different forms which makes segmentation a difficult and complex task. Different methods are present in literature for improving results for skin lesion segmentation. Each method has some pros and cons and it is observed that none of them can be regarded as a generalized method working for all types of skin lesions. The paper proposes an algorithm that combines the advantages of clustering, thresholding and active contour methods currently being used independently for segmentation purposes. A modified algorithm for thresholding based on fusion of Fuzzy C mean clustering and histogram thresholding is applied to initialize level set automatically and also for estimating controlling parameters for level set evolution. The performance of level set segmentation is subject to appropriate initialization, so the proposed algorithm is being compared with some other state-of-the-art initialization methods. The work has been tested on clinical database of 270 images. Parameters for performance evaluation are presented in detail. Increased true detection rate and reduced false positive and false negative errors confirm the effectiveness of the proposed method for skin cancer detection.
收起
摘要 :
In this paper, we present an incremental learning method on a budget for embedded systems. We discuss its application for two power systems: a micro-converter for photovoltaic and a step down DCDC- converter. This learning method ...
展开
In this paper, we present an incremental learning method on a budget for embedded systems. We discuss its application for two power systems: a micro-converter for photovoltaic and a step down DCDC- converter. This learning method is a variation of the general regression neural network but it is able to continue incremental learning on a bounded support set. The method basically learns new instances by adding new kernels. However, when the number of kernels reaches a predefined upper bound, the method selects the most effective learning option from several options: including replacing the most ineffective kernel with the new kernel, modifying of the parameters of existing kernels, and ignoring the new instance. The proposed method is compared with other similar learning methods on a budget, which are based on kernel perceptron. Two examples of the application of the proposed method are demonstrated in power electronics. In these two examples, we show that the proposed system learns the properties of the control-objects during the services and realizes quick control.
收起
摘要 :
In this paper we propose a new local learning algorithm for appearance-based object pose estimation, called Locally Linearly Embedded Regression (LLER). LLER uses a constrained version of Locally Linear Embedding (LLE) to simultan...
展开
In this paper we propose a new local learning algorithm for appearance-based object pose estimation, called Locally Linearly Embedded Regression (LLER). LLER uses a constrained version of Locally Linear Embedding (LLE) to simultaneously embed into an intermediate low-dimensional space the training images, the query image and a grid of pose parameters. A linear map is learned between the points in the local neighborhood of the query representation in this low-dimensional intermediate space and their corresponding pose parameters, which is used to directly recover the pose of the query image. The proposed method has been evaluated in a pose estimation task on a database of 16 different objects, consistently outperforming several representative global and local appearance-based pose estimation methods.
收起
摘要 :
Researchers have been building robots able to interact and work with people at home. To share and reuse robot code between different developers, we present a service-based approach that exploits the standard web interface to creat...
展开
Researchers have been building robots able to interact and work with people at home. To share and reuse robot code between different developers, we present a service-based approach that exploits the standard web interface to create reusable robotic services. Our approach includes knowledge ontology planning and neural network learning strategies for robot control. In addition, several service functions, including service discovery, selection, composition, and reconfiguration have been developed for operating the services. The proposed approach has been implemented and evaluated. The results show that our approach can be used to build robotic services successfully.
收起
摘要 :
This paper presents an alternative approach for the control and balancing operations of an inverted pendulum. The proposed method uses a neuronal network called NeuraBase to learn the sensor events obtained via a rotary encoder an...
展开
This paper presents an alternative approach for the control and balancing operations of an inverted pendulum. The proposed method uses a neuronal network called NeuraBase to learn the sensor events obtained via a rotary encoder and the motor events controlling a stepper motor, which rotates the swinging arm. A neuron layer called the controller network will link the sensor neuron events to the motor neurons. The proposed NeuraBase network model (NNM) has demonstrated its ability to successfully control the balancing operation of the pendulum in the absence of a dynamic model and theoretical control methods. The controller also demonstrated its robustness in the adaptive learning of pendulum balancing with imposed system changes.
收起
摘要 :
Emotional semantic image retrieval systems aim at incorporating the user's affective states for responding adequately to the user's interests. One challenge is to select features specific to image affect detection. Another challen...
展开
Emotional semantic image retrieval systems aim at incorporating the user's affective states for responding adequately to the user's interests. One challenge is to select features specific to image affect detection. Another challenge is to build effective learning models or classifiers to bridge the so-called "affective gap". In this work, we study the affective classification and retrieval of abstract images by applying multiple kernel learning framework. An image can be represented by different feature spaces and multiple kernel learning can utilize all these feature representations simultaneously (i.e., multiview learning), such that it jointly learns the feature representation weights and corresponding classifier in an intelligent manner. Our experimental results on two abstract image datasets demonstrate the advantage of the multiple kernel learning framework for image affect detection in terms of feature selection, classification performance, and interpretation.
收起
摘要 :
The paper describes a unified algorithm for both parametric and structural identification. The approach combines three typical techniques such as neural networks, statistics and genetic algorithm. A specific structure of the neura...
展开
The paper describes a unified algorithm for both parametric and structural identification. The approach combines three typical techniques such as neural networks, statistics and genetic algorithm. A specific structure of the neural network is used that allows to design a controller directly from parameters of the identified model. The control strategy based on reference model is discussed. Finally, the proposed solution is illustrated by numerical example.
收起
摘要 :
In this paper, we address a knee joint orthosis control for rehabilitation purposes. Only the structure of the system's dynamic model is supposed to be known. Inertia of the knee-shank-orthosis system is identified on-line using a...
展开
In this paper, we address a knee joint orthosis control for rehabilitation purposes. Only the structure of the system's dynamic model is supposed to be known. Inertia of the knee-shank-orthosis system is identified on-line using an adaptive term. In order to approximate all of the other dynamics (viscous and solid frictions, gravity related torque, etc.), we use an RBF Neural Network (RBFNN) with no off-line prior training. Adaptation laws of the neural parameters and the inertia adaptive term are derived from the closed loop system's overall stability study using Lyapunov's theory. The study considers three cases: wearer being completely inactive or applying either a resistive or an assistive torque. Simulation results and conducted analysis show the effectiveness of the proposed approach.
收起
摘要 :
Multi-label support vector machine (Rank-SVM) is an effective algorithm for multi-label classification, which is formulated as a quadratic programming problem with q equality constraints and lots of box constraints for a q-class m...
展开
Multi-label support vector machine (Rank-SVM) is an effective algorithm for multi-label classification, which is formulated as a quadratic programming problem with q equality constraints and lots of box constraints for a q-class multi-label data set. So far, Rank-SVM is solved by Frank-Wolfe method (FWM), where a large-scale linear programming problem needs to be dealt with at each iteration. In this paper, we propose a random block coordinate descent method (RBCDM) for Rank-SVM, in which a small-scale quadratic programming problem with at least (q+1) variables randomly is solved at each iteration. Experiments on three data sets illustrate that our RBCDM runs much faster than FWM for Rank-SVM, and Rank-SVM is a powerful candidate for multi-label classification.
收起