İlyas Çankaya

Performance evaluation of prototype-based clustering algorithms combined MDL index

Clustering algorithms are used to group data depending on a distance. Best clustering analysis should be resisting the presence of outliers, less sensitive to initialization as well as the input sequence ordering. This article compares the performance among three of prototype‐based unsupervised clustering algorithms: Neural Gas (NG), Growing Neural Gas (GNG) and Robust Growing Neural Gas (RGNG). Based on NG and GNG, there are different clustering algorithms proposed and suggested in different literatures. So, in this work a comparison between the two basic clustering algorithms NG and GNG have presented using the performance evaluation of these techniques, in contrast to the RGNG which was proposed within the GNG. Another comparison due to the MDL criterion between RGNG that used MDL value as the clustering validity index, versus GNG and NG combined with MDL. Statistical estimations are applied to explain the meaning of the output results when these algorithms fed to the synthetic 2D dataset. Moreover, a simple software package is designed and implemented as an automatic clustering model for any dataset to use as a part of the neural network course. NG, GNG and RGNG algorithms are performed in the same package using a MATLAB‐based Graphical User Interface (GUI) tool. This visual tool lets the students/ researchers visualize the desired results using plots also clicking a few buttons.

Clustering fMRI data with a robust unsupervised learning algorithm for neuroscience data mining

BACKGROUND Clustering approaches used in functional magnetic resonance imaging (fMRI) research use brain activity to divide the brain into various parcels with some degree of homogeneous characteristics, but choosing the appropriate clustering algorithms remains a problem. NEW METHOD A novel application of the robust unsupervised learning approach is proposed in the current study. Robust growing neural gas (RGNG) algorithm was fed into fMRI data and compared with growing neural gas (GNG) algorithm, which has not been used for this purpose or any other medical application. Learning algorithms proposed in the current study are fed with real and free auditory fMRI datasets. RESULTS The fMRI result obtained by running RGNG was within the expected outcome and is similar to those found with the hypothesis method in detecting active areas within the expected auditory cortices. COMPARISON WITH EXISTING METHOD(S) The fMRI application of the presented RGNG approach is clearly superior to other approaches in terms of its insensitivity to different initializations and the presence of outliers, as well as its ability to determine the actual number of clusters successfully, as indicated by its performance measured by minimum description length (MDL) and receiver operating characteristic (ROC) analysis. CONCLUSIONS The RGNG can detect the active zones in the brain, analyze brain function, and determine the optimal number of underlying clusters in fMRI datasets. This algorithm can define the positions of the center of an output cluster corresponding to the minimal MDL value.

Sparse fast Fourier transform for exactly sparse signals and signals with additive Gaussian noise

In recent years, the Fourier domain representation of sparse signals has been very attractive. Sparse fast Fourier transform (or sparse FFT) is a new technique which computes the Fourier transform in a compressed way, using only a subset of the input data. Sparse FFT computes the desired transform in sublinear time, which means in an amount of time that is smaller than the size of data. In big data problems and medical imaging to reduce the time that patient spends in MRI machine, FFT algorithm is not ‘fast’ enough anymore; therefore, the concept of sparse FFT is very important. Similar to compressed sensing, sparse FFT algorithm computes just the important components in the frequency domain in sublinear time. In this work, sparse FFT algorithm is studied and implemented on MATLAB and its performance is compared with Ann Arbor FFT. A filter is used to hash the frequencies in the n dimensional frequency-sparse signal into B bins, where

A method for the Computational Frequency Sweep Analysis of Nonlinear ODEs using GPU Acceleration

B=n/16

Super-resolution image reconstruction applied to an active millimeter wave imaging system based on compressive sensing

B=n/16. The filter is used for analyzing an important fraction of the whole signal, and therefore, instead of computing n point FFT, B point FFT is computed, and this results in a faster Fourier transform. The probability of success of the implemented algorithm is investigated for noiseless and noisy signals. It is deduced that as the sparsity increases, the probability of perfect transform also increases. If the performances of the algorithm in both cases are compared, it is clearly seen that the performances degrade when there is noise. Therefore, it can be concluded that the algorithm should be improved especially for noisy considerations. The solvability boundary for a constant probability of error is deducted and added to give insight for future studies.

MATLAB VE ASP.NET TABANLI WEB ARAYÜZÜ KULLANILARAK DOĞRUSAL OLMAYAN SİSTEMLERİN ANALİZİ

Electrocardiogram (ECG) signals are continuously monitored for early diagnosis of heart diseases. However, a long-term monitoring generates large amounts of data at a level that makes storage and transmission difficult. Moreover, these records may be subject to different types of noise distributions resulting from operating conditions. Therefore, an effective and reliable data compression technique is needed for ECG data transmission, storage and analysis without losing the clinical information content. This study proposes the e-insensitive Huber loss based support vector regression for the compressing of ECG signals. Since the Huber loss function is a mixture of quadratic and linear loss functions, it can properly take into account the different noise types in the data set. Compression performance of the proposed method has been assessed using ECG records from the MIT-BIH arrhythmia database. Experimental results demonstrate that the proposed loss function is an attractive candidate for compressing ECG data.

BiLGiSAYAR KONTROLLÜ SERBEST DÜŞME DENEY SİSTEMİNİN TASARIMI

ABSTRACT Computational sweep analysis of nonlinear ODEs (ordinary differential equations) is of importance in engineering system analysis and design. Sweep analyses usually demand intense computational power according to the number of points and the number of system parameters. This paper presents an efficient parallel algorithm for the sweep analysis of nonlinear ODEs based on graphical processing unit acceleration. The developed method preserves the jump phenomenon characteristics intrinsic to nonlinear ODEs and reduces the effects of irregular computational load. Experiments were realized using Duffing equation by sweeping frequency, amplitude, and equation coefficients. Directly, data parallel implementation and proposed implementations are compared to show the efficiency of the proposed method. Experimental results show that the new method provides significant reductions in the computational durations when compared to sequential implementation.

Broken rotor bar fault detection in inverter-fed squirrel cage induction motors using stator current analysis and fuzzy logic

Compressed Sensing (CS) with Cyclic-S Hadamard matrix is proposed for single pixel imaging applications in this study. In single pixel imaging scheme, N = r · c samples should be taken for r×c pixel image where · denotes multiplication. CS is a popular technique claiming that the sparse signals can be reconstructed with samples under Nyquist rate. Therefore to solve the slow data acquisition problem in Terahertz (THz) single pixel imaging, CS is a good candidate. However, changing mask for each measurement is a challenging problem since there is no commercial Spatial Light Modulators (SLM) for THz band yet, therefore circular masks are suggested so that for each measurement one or two column shifting will be enough to change the mask. The CS masks are designed using cyclic-S matrices based on Hadamard transform for 9 × 7 and 15 × 17 pixel images within the framework of this study. The %50 compressed images are reconstructed using total variation based TVAL3 algorithm. Matlab simulations demonstrates that ...