Moongu Jeon

Structure Preserving Dimension Reduction for Clustered Text Data Based on the Generalized Singular Value Decomposition

In todays vector space information retrieval systems, dimension reduction is imperative for efficiently manipulating the massive quantity of data. To be useful, this lower-dimensional representation must be a good approximation of the full document set. To that end, we adapt and extend the discriminant analysis projection used in pattern recognition. This projection preserves cluster structure by maximizing the scatter between clusters while minimizing the scatter within clusters. A common limitation of trace optimization in discriminant analysis is that one of the scatter matrices must be nonsingular, which restricts its application to document sets in which the number of terms does not exceed the number of documents. We show that by using the generalized singular value decomposition (GSVD), we can achieve the same goal regardless of the relative dimensions of the term-document matrix. In addition, applying the GSVD allows us to avoid the explicit formation of the scatter matrices in favor of working directly with the data matrix, thus improving the numerical properties of the approach. Finally, we present experimental results that confirm the effectiveness of our approach.

Despeckling of medical ultrasound images using Daubechies complex wavelet transform

The paper presents a novel despeckling method, based on Daubechies complex wavelet transform, for medical ultrasound images. Daubechies complex wavelet transform is used due to its approximate shift invariance property and extra information in imaginary plane of complex wavelet domain when compared to real wavelet domain. A wavelet shrinkage factor has been derived to estimate the noise-free wavelet coefficients. The proposed method firstly detects strong edges using imaginary component of complex scaling coefficients and then applies shrinkage on magnitude of complex wavelet coefficients in the wavelet domain at non-edge points. The proposed shrinkage depends on the statistical parameters of complex wavelet coefficients of noisy image which makes it adaptive in nature. Effectiveness of the proposed method is compared on the basis of signal to mean square error (SMSE) and signal to noise ratio (SNR). The experimental results demonstrate that the proposed method outperforms other conventional despeckling methods as well as wavelet based log transformed and non-log transformed methods on test images. Application of the proposed method on real diagnostic ultrasound images has shown a clear improvement over other methods.

Neuromorphic Hardware System for Visual Pattern Recognition With Memristor Array and CMOS Neuron

This paper presents a neuromorphic system for visual pattern recognition realized in hardware. A new learning rule based on modified spike-timing-dependent plasticity is also presented and implemented with passive synaptic devices. The system includes an artificial photoreceptor, a Pr0.7Ca0.3MnO3-based memristor array, and CMOS neurons. The artificial photoreceptor consisting of a CMOS image sensor and a field-programmable gate array converts an image into spike signals, and the memristor array is used to adjust the synaptic weights between the input and output neurons according to the learning rule. A leaky integrate-and-fire model is used for the output neuron that is built together with the image sensor on a single chip. The system has 30 input neurons that are interconnected to 10 output neurons through 300 memristors. Each input neuron corresponding to a pixel in a 5 × 6 pixel image generates voltage pulses according to the pixel value. The voltage pulses are then weighted and integrated by the memristors and the output neurons, respectively, to be compared with a certain threshold voltage above which an output neuron fires. The system has been successfully demonstrated by training and recognizing number images from 0 to 9.

RRAM-based synapse for neuromorphic system with pattern recognition function

Feasibility of a high speed pattern recognition system using 1k-bit cross-point synaptic RRAM array and CMOS-based neuron chip has been experimentally demonstrated. Learning capability of a neuromorphic system comprising RRAM synapses and CMOS neurons has been confirmed experimentally, for the first time.

Electronic system with memristive synapses for pattern recognition

Memristive synapses, the most promising passive devices for synaptic interconnections in artificial neural networks, are the driving force behind recent research on hardware neural networks. Despite significant efforts to utilize memristive synapses, progress to date has only shown the possibility of building a neural network system that can classify simple image patterns. In this article, we report a high-density cross-point memristive synapse array with improved synaptic characteristics. The proposed PCMO-based memristive synapse exhibits the necessary gradual and symmetrical conductance changes, and has been successfully adapted to a neural network system. The system learns, and later recognizes, the human thought pattern corresponding to three vowels, i.e. /a /, /i /, and /u/, using electroencephalography signals generated while a subject imagines speaking vowels. Our successful demonstration of a neural network system for EEG pattern recognition is likely to intrigue many researchers and stimulate a new research direction.

Nanoscale RRAM-based synaptic electronics: toward a neuromorphic computing device

Efforts to develop scalable learning algorithms for implementation of networks of spiking neurons in silicon have been hindered by the considerable footprints of learning circuits, which grow as the number of synapses increases. Recent developments in nanotechnologies provide an extremely compact device with low-power consumption.In particular, nanoscale resistive switching devices (resistive random-access memory (RRAM)) are regarded as a promising solution for implementation of biological synapses due to their nanoscale dimensions, capacity to store multiple bits and the low energy required to operate distinct states. In this paper, we report the fabrication, modeling and implementation of nanoscale RRAM with multi-level storage capability for an electronic synapse device. In addition, we first experimentally demonstrate the learning capabilities and predictable performance by a neuromorphic circuit composed of a nanoscale 1 kbit RRAM cross-point array of synapses and complementary metal-oxide-semiconductor neuron circuits. These developments open up possibilities for the development of ubiquitous ultra-dense, ultra-low-power cognitive computers.

Neuromorphic speech systems using advanced ReRAM-based synapse

We demonstrate an advanced ReRAM based analog artificial synapse for neuromorphic systems. Nitrogen doped TiN/PCMO based artificial synapse is proposed to improve the performance and reliability of the neuromorphic systems by using simple identical spikes. For the first time, we develop fully unsupervised learning with proposed analog synapses which is illustrated with the help of auditory and electroencephalography (EEG) applications.

Statistical properties analysis of real world tournament selection in genetic algorithms

Abstract Genetic algorithms (GAs) are probabilistic optimization methods based on the biological principle of natural evolution. One of the important operators in GAs is the selection strategy for obtaining better solutions. Specifically, finding a balance between the selection pressure and diversity is a critical issue in designing an efficient selection strategy. To this extent, the recently proposed real world tournament selection (RWTS) method has showed good performance in various benchmark problems. In this paper, we focus on analyzing characteristics of RWTS from the viewpoint of both the selection probabilities and stochastic sampling properties in order to provide a rational explanation for why RWTS provides improved performance. Statistical experimental results show that RWTS has a higher selection pressure with a relatively small loss of diversity and higher sampling accuracy than conventional tournament selection. The performance tests in a traveling salesman problem further confirm that the comparatively higher pressure and sampling accuracy, which are inherent in RWTS, can enhance the performance in the selection strategy.

Anatomical evaluation of CT-MRI combined femoral model.

BackgroundBoth CT and MRI are complementary to each other in that CT can produce a distinct contour of bones, and MRI can show the shape of both ligaments and bones. It will be ideal to build a CT-MRI combined model to take advantage of complementary information of each modality. This study evaluated the accuracy of the combined femoral model in terms of anatomical inspection.MethodsSix normal porcine femora (180 ± 10 days, 3 lefts and 3 rights) with ball markers were scanned by CT and MRI. The 3D/3D registration was performed by two methods, i.e. the landmark-based 3 points-to-3 points and the surface matching using the iterative closest point (ICP) algorithm. The matching accuracy of the combined model was evaluated with statistical global deviation and locally measure anatomical contour-based deviation. Statistical analysis to assess any significant difference between accuracies of those two methods was performed using univariate repeated measures ANOVA with the Turkey post hoc test.ResultsThis study revealed that the local 2D contour-based measurement of matching deviation was 0.5 ± 0.3 mm in the femoral condyle, and in the middle femoral shaft. The global 3D contour matching deviation of the landmark-based matching was 1.1 ± 0.3 mm, but local 2D contour deviation through anatomical inspection was much larger as much as 3.0 ± 1.8 mm.ConclusionEven with human-factor derived errors accumulated from segmentation of MRI images, and limited image quality, the matching accuracy of CT-&-MRI combined 3D models was 0.5 ± 0.3 mm in terms of local anatomical inspection.

A new framework for background subtraction using multiple cues

In this work, to effectively detect moving objects in a fixed camera scene, we propose a novel background subtraction framework employing diverse cues: pixel texture, pixel color and region appearance. The texture information of the scene is clustered by the conventional codebook based background modeling technique, and utilized to detect initial foreground regions. In this process, we employ a new texture operator namely, scene adaptive local binary pattern (SALBP) that provides more consistent and accurate texture-code generation by applying scene adaptive multiple thresholds. Background statistics of the color cues are also modeled by the codebook scheme and employed to refine the texture-based detection results by integrating color and texture characteristics. Finally, appearance of each refined foreground blob is verified by measuring the partial directed Hausdorff distance between the shape of a blob boundary and the edge-map of the corresponding sub-image region in the input frame. The proposed method is compared with other state-of-the-art background subtraction techniques and its results demonstrate that our method outperforms others for complicated environments in video surveillance applications.