Browsing by Subject "Artificial neural networks"

Now showing 1 - 20 of 23

Open Access
Activity recognition invariant to sensor orientation with wearable motion sensors
(MDPI AG, 2017) Yurtman, A.; Barshan, B.
Most activity recognition studies that employ wearable sensors assume that the sensors are attached at pre-determined positions and orientations that do not change over time. Since this is not the case in practice, it is of interest to develop wearable systems that operate invariantly to sensor position and orientation. We focus on invariance to sensor orientation and develop two alternative transformations to remove the effect of absolute sensor orientation from the raw sensor data. We test the proposed methodology in activity recognition with four state-of-the-art classifiers using five publicly available datasets containing various types of human activities acquired by different sensor configurations. While the ordinary activity recognition system cannot handle incorrectly oriented sensors, the proposed transformations allow the sensors to be worn at any orientation at a given position on the body, and achieve nearly the same activity recognition performance as the ordinary system for which the sensor units are not rotatable. The proposed techniques can be applied to existing wearable systems without much effort, by simply transforming the time-domain sensor data at the pre-processing stage. © 2017 by the authors. Licensee MDPI, Basel, Switzerland.
Open Access
An approach based on sound classification to predict soundscape perception through machine learning
(2021-06) Acun, Volkan
A growing amount of literature and a series of ISO standards focus on concept, data collection, and data analysis methods of soundscapes. Yet, this field of research still lacks predictive models. We hypothesize that machine learning methods can be used to develop a predictive model by identifying the audio content of soundscapes and correlating it with individuals’ perceived affective response to the soundscapes. Therefore, this research aims to identify machine learning-based sound classification methods for analyzing the audio content of soundscapes and using its output in a second model for evaluating the association between the audio content and perception of the soundscape. We focused on museum soundscapes to conduct our research. The methodology of this thesis is divided into two parts. For the first part, we used Convolutional Neural Networks for classifying the audio content of the soundscape. Due to their limitations, we used a different approach rather than the typical environmental sound classification methods. We used musical instruments for the training dataset and optimized the neural network for this type of task. The convolutional neural network classified the audio content of the soundscapes based on their similarities to the musical instruments of the dataset. We conducted an online soundscape perception survey to measure participants' affective responses to different museum soundscapes for the second part. To predict individuals’ perception of soundscapes, we developed a feedforward neural network model. This model used the audio content output from the sound classification model and the soundscape survey data to predict the perceived affective quality of soundscapes. We concluded the thesis by conducting statistical analyses to explore the association between the variable used in the predictive model.
Open Access
Classifying human leg motions with uniaxial piezoelectric gyroscopes
(2009) Tunçel O.; Altun, K.; Barshan, B.
This paper provides a comparative study on the different techniques of classifying human leg motions that are performed using two low-cost uniaxial piezoelectric gyroscopes worn on the leg. A number of feature sets, extracted from the raw inertial sensor data in different ways, are used in the classification process. The classification techniques implemented and compared in this study are: Bayesian decision making (BDM), a rule-based algorithm (RBA) or decision tree, least-squares method (LSM), k-nearest neighbor algorithm (k-NN), dynamic time warping (DTW), support vector machines (SVM), and artificial neural networks (ANN). A performance comparison of these classification techniques is provided in terms of their correct differentiation rates, confusion matrices, computational cost, and training and storage requirements. Three different cross-validation techniques are employed to validate the classifiers. The results indicate that BDM, in general, results in the highest correct classification rate with relatively small computational cost. © 2009 by the authors.
Open Access
Comparative analysis of different approaches to target classification and localization with sonar
(IEEE, 2001-08) Ayrulu, Birsel; Barshan, Billur
The comparison of different classification and fusion techniques was done for target classification and localization with sonar. Target localization performance of artificial neural networks (ANN) was found to be better than the target differentiation algorithm (TDA) and fusion techniques. The target classification performance of non-parametric approaches was better than that of parameterized density estimator (PDE) using homoscedastic and heteroscedastic NM for statistical pattern recognition techniques.
Open Access
A comparative analysis of different approaches to target differentiation and localization using infrared sensors
(2006) Aytaç, Tayfun
This study compares the performances of various techniques for the differentiation and localization of commonly encountered features in indoor environments, such as planes, corners, edges, and cylinders, possibly with different surface properties, using simple infrared sensors. The intensity measurements obtained from such sensors are highly dependent on the location, geometry, and surface properties of the reflecting feature in a way that cannot be represented by a simple analytical relationship, therefore complicating the localization and differentiation process. The techniques considered include rule-based, template-based, and neural network-based target differentiation, parametric surface differentiation, and statistical pattern recognition techniques such as parametric density estimation, various linear and quadratic classifiers, mixture of normals, kernel estimator, k-nearest neighbor, artificial neural network, and support vector machine classi- fiers. The geometrical properties of the targets are more distinctive than their surface properties, and surface recognition is the limiting factor in differentiation. Mixture of normals classifier with three components correctly differentiates three types of geometries with different surface properties, resulting in the best performance (100%) in geometry differentiation. For a set of six surfaces, we get a correct differentiation rate of 100% in parametric differentiation based on reflection modeling. The results demonstrate that simple infrared sensors, when coupled with appropriate processing, can be used to extract substantially more information than such devices are commonly employed for. The demonstrated system would find application in intelligent autonomous systems such as mobile robots whose task involves surveying an unknown environment made of different geometry and surface types. Industrial applications where different materials/surfaces must be identified and separated may also benefit from this approach.
Open Access
Comparative analysis of different approaches to target differentiation and localization with sonar
(Elsevier, 2003) Barshan, B.; Ayrulu, B.
This study compares the performances of different methods for the differentiation and localization of commonly encountered features in indoor environments. Differentiation of such features is of interest for intelligent systems in a variety of applications such as system control based on acoustic signal detection and identification, map building, navigation, obstacle avoidance, and target tracking. Different representations of amplitude and time-of-2ight measurement patterns experimentally acquired from a real sonar system are processed. The approaches compared in this study include the target differentiation algorithm, Dempster-Shafer evidential reasoning, different kinds of voting schemes, statistical pattern recognition techniques (k-nearest neighbor classifier, kernel estimator, parameterized density estimator, linear discriminant analysis, and fuzzy c-means clustering algorithm), and artificial neural networks. The neural networks are trained with different input signal representations obtained usingpre-processing techniques such as discrete ordinary and fractional Fourier, Hartley and wavelet transforms, and Kohonen's self-organizing feature map. The use of neural networks trained with the back-propagation algorithm, usually with fractional Fourier transform or wavelet pre-processing results in near perfect differentiation, around 85% correct range estimation and around 95% correct azimuth estimation, which would be satisfactory in a wide range of applications. © 2002 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.
Open Access
Comparative study on classifying human activities with miniature inertial and magnetic sensors
(Elsevier, 2010) Altun, K.; Barshan, B.; Tunçel, O.
This paper provides a comparative study on the different techniques of classifying human activities that are performed using body-worn miniature inertial and magnetic sensors. The classification techniques implemented and compared in this study are: Bayesian decision making (BDM), a rule-based algorithm (RBA) or decision tree, the least-squares method (LSM), the k-nearest neighbor algorithm (k-NN), dynamic time warping (DTW), support vector machines (SVM), and artificial neural networks (ANN). Human activities are classified using five sensor units worn on the chest, the arms, and the legs. Each sensor unit comprises a tri-axial gyroscope, a tri-axial accelerometer, and a tri-axial magnetometer. A feature set extracted from the raw sensor data using principal component analysis (PCA) is used in the classification process. A performance comparison of the classification techniques is provided in terms of their correct differentiation rates, confusion matrices, and computational cost, as well as their pre-processing, training, and storage requirements. Three different cross-validation techniques are employed to validate the classifiers. The results indicate that in general, BDM results in the highest correct classification rate with relatively small computational cost.
Open Access
Deep learning with extended exponential linear unit (DELU)
(Springer, 2023-08-16) Çatalbaş, Burak; Morgül, Ömer
Activation functions are crucial parts of artificial neural networks. From the first perceptron created artificially up to today, many functions are proposed. Some of them are currently in common use, such as Rectified Linear Unit (ReLU) and Exponential Linear Unit (ELU) and other ReLU variants. In this article we propose a novel activation function, called ExtendeD Exponential Linear Unit (DELU). After its introduction and presenting its basic properties, by making various simulations with different datasets and architectures, we show that it may perform better than other activation functions in certain cases. While also inheriting most of the good properties of ReLU and ELU, DELU offers an increase of success in comparison with them by slowing the alignment of neurons in early stages of training process. In experiments, DELU performed better than other activation functions in general, for Fashion MNIST, CIFAR-10 and CIFAR-100 classification tasks with different sized Residual Neural Networks (ResNet). Specifically, DELU managed to reduce the error rate by sufficiently high confidence levels in CIFAR datasets in comparison with ReLU and ELU networks. In addition, DELU is compared in an image segmentation example as well. Also, compatibility of DELU is tested with different initializations, and statistical methods are employed to verify these success rates by using Z-score analysis, which may be considered as a different view of success assessment in neural networks.
Open Access
Farklı yapay sinir ağı temelli sınıflandırıcılar ile insan hareketi tanımlama
(IEEE, 2017-05) Çatalbaş, Burak; Morgül, Ömer; Çatalbaş, Bahadır
İnsan Hareketi Tanımlanması, taşıdığı önem ve sınırlı öznitelik vektörü ile yüksek sınıflandırma oranlarına ulaşmasında karşılaşılan zorluk nedeniyle popüler bir araştırma konusudur. Bireylerin hareket ölçülebilirliginin akıllı telefonların içinde gömülü bulunan atalet ölçüm birimleri sayesinde artması ile birlikte, bu alanda toplanan veri miktarı artmakta ve daha başarılı sınıflandırıcıların tasarlanabilmesine imkan saglanmaktadır. Yapay sinir ağları, konvansiyonel sınıflandırıcılara göre sınıflandırma sorunlarında daha iyi performans sergileyebilmektedir. Bu çalışmada, Irvine Kaliforniya Üniversitesi (UCI) veri setine yapay sinir ağı temelli bir sınıflandırıcı önermek için çeşitli yapay sinir ağı yapıları denenmiş olup, bu sınıflandırıcılar ile elde edilen başarı oranları literatürdeki aynı veri kümesi için bulunan sonuçlarla karşılaştırılmıştır.
Embargo
Identification and adaptive control of bipedal robot motion with artificial neural networks
(2024-07) Çatalbaş, Burak
Artificial Neural Networks (ANNs) is one of the most popular fields of machine learning thanks to the critical improvements in the last decade, including their applications in the field of robotics and control. The important usage of neural networks in robotics makes it possible for robots to act and interact similar to humans. In this manner, legged robots are important platforms to mimick human locomotion. However, there are significant difficulties to apply system identification and control schemes for these hybrid dynamical structures. With this purpose, this thesis focuses on using artificial neural network-based novel techniques on these problems, for reaching to an efficient walking ability for bipedal robot systems like their counterparts in the nature. In this thesis, our work to find and apply our novel techniques is mainly divided into two parts. In the first part, inspired by a class of activation functions frequently used in deep learning literature, we propose a novel activation function and investigate its performance in various segmentation and classification tasks by using different well-known datasets. In the second part of the thesis, biped robot locomotion is chosen as the main topic. Separate datasets are created for three experiment configurations. For 2D and 3D simulations, locomotion control, system identification and adaptive control are applied with neural networks for successful periodical walking with low errors, having approximations of robot models and preparing for the adaptive learning using both control and identification blocks, respectively. For 2D physical robot system, system identification is completed for a walking dataset generated with varying speed levels. For all cases, proposed novel activation function DELU (ExtendeD Exponential Linear Unit) and its tuned functions are tried together with other activation functions in comparison, to reach better performances.
Open Access
Indoor localization with transfer learning
(IEEE, 2022-08-29) Korkmaz, İlter Onat; Özateş, Tuna; Koç, Enes; Aydın, Ege; Kor, Ege; Dilek, Doğaç; Güngen, Murat Alp; Köse, İdil Gökalp; Akman, Çağlar
Indoor positioning methods aim to estimate positions of transmitters where the GPS signals are unavailable. These systems usually employ algorithms explicitly trained for a single location such as fingerprinting method. For that reason, they can only be used in a particular location. This restriction prevents the use of the fingerprint method in tasks such as search and rescue operations where there is no prior knowledge of the place. A fingerprinting system using a trained algorithm with data collected from many places can work in multiple places. This paper proposes an indoor positioning system that uses the parameters of a pre-trained neural network trained with the data obtained from finite difference time domain simulations with transfer learning without collecting large amounts of data. The initial parameters for the model to be trained with the received signal strength (RSS) data collected from real places are used as be the parameters of the artificial neural network trained with the aforementioned simulation data. Performance results of the trained model are comparable to the results of the works in which fingerprinting method is employed in a single environment.
Open Access
Leg motion classification with artificial neural networks using wavelet-based features of gyroscope signals
(2011) Ayrulu-Erdem, B.; Barshan, B.
We extract the informative features of gyroscope signals using the discrete wavelet transform (DWT) decomposition and provide them as input to multi-layer feed-forward artificial neural networks (ANNs) for leg motion classification. Since the DWT is based on correlating the analyzed signal with a prototype wavelet function, selection of the wavelet type can influence the performance of wavelet-based applications significantly. We also investigate the effect of selecting different wavelet families on classification accuracy and ANN complexity and provide a comparison between them. The maximum classification accuracy of 97.7% is achieved with the Daubechies wavelet of order 16 and the reverse bi-orthogonal (RBO) wavelet of order 3.1, both with similar ANN complexity. However, the RBO 3.1 wavelet is preferable because of its lower computational complexity in the DWT decomposition and reconstruction. © 2011 by the authors; licensee MDPI, Basel, Switzerland.
Open Access
Micro-Architectural features as soft-error markers in embedded safety-critical systems: preliminary study
(Institute of Electrical and Electronics Engineers Inc., 2023-07-12) Kasap, Deniz; Carpegna, A.; Savino, A.; Di Carlo, S.
Radiation-induced soft errors are one of the most challenging issues in Safety Critical Real-Time Embedded System (SACRES) reliability, usually handled using different flavors of Double Modular Redundancy (DMR) techniques. This solution is becoming unaffordable due to the complexity of modern micro-processors in all domains. This paper addresses the promising field of using Artificial Intelligence (AI) based hardware detectors for soft errors. To create such cores and make them general enough to work with different software applications, micro-Architectural attributes are a fascinating option as candidate fault detection features. Several processors already track these features through dedicated Performance Monitoring Unit (PMU). However, there is an open question to understand to what extent they are enough to detect faulty executions. Exploiting the capability of gem5 to simulate real computing systems, perform fault injection experiments, and profile micro-Architectural attributes (i.e., gem5 Stats), this paper presents the results of a comprehensive analysis regarding the potential attributes to detect soft errors and the associated models that can be trained with these features.
Open Access
Neural network-based target differentiation using sonar for robotics applications
(IEEE, 2000-08) Barshan, B.; Ayrulu, B.; Utete, S. W.
This study investigates the processing of sonar signals using neural networks for robust differentiation of commonly encountered features in indoor robot environments. The neural network can differentiate more targets with higher accuracy, improving on previously reported methods. It achieves this by exploiting the identifying features in the differential amplitude and time-of-flight (TOF) characteristics of these targets. Robustness tests indicate that the amplitude information is more crucial than TOF for reliable operation. The study suggests wider use of neural networks and amplitude information in sonar-based mobile robotics.
Open Access
Neural networks for improved target differentiation and localization with sonar
(Pergamon Press, 2001) Ayrulu, B.; Barshan, B.
This study investigates the processing of sonar signals using neural networks for robust differentiation of commonly encountered features in indoor robot environments. Differentiation of such features is of interest for intelligent systems in a variety of applications. Different representations of amplitude and time-of-flight measurement patterns acquired from a real sonar system are processed. In most cases, best results are obtained with the low-frequency component of the discrete wavelet transform of these patterns. Modular and non-modular neural network structures trained with the back-propagation and generating-shrinking algorithms are used to incorporate learning in the identification of parameter relations for target primitives. Networks trained with the generating-shrinking algorithm demonstrate better generalization and interpolation capability and faster convergence rate. Neural networks can differentiate more targets employing only a single sensor node, with a higher correct differentiation percentage (99%) than achieved with previously reported methods (61-90%) employing multiple sensor nodes. A sensor node is a pair of transducers with fixed separation, that can rotate and scan the target to collect data. Had the number of sensing nodes been reduced in the other methods, their performance would have been even worse. The success of the neural network approach shows that the sonar signals do contain sufficient information to differentiate all target types, but the previously reported methods are unable to resolve this identifying information. This work can find application in areas where recognition of patterns hidden in sonar signals is required. Some examples are system control based on acoustic signal detection and identification, map building, navigation, obstacle avoidance, and target-tracking applications for mobile robots and other intelligent systems. Copyright © 2001 Elsevier Science Ltd.
Open Access
Neural-network quantum states for a two-leg bose-hubbard ladder under a synthetic magnetic field
(2023-07) Çeven, Kadir
This thesis explores novel quantum phases in a two-leg Bose-Hubbard ladder, achieved using neural-network quantum states. The remarkable potential of quantum gas systems for analog quantum simulation of strongly correlated quantum matter is well-known; however, it is equally evident that new theoretical bases are urgently required to comprehend their intricacies fully. While simple one dimensional models have served as valuable test cases, ladder models naturally emerge as the next step, enabling studying higher dimensional effects, including gauge fields. Utilizing the paper [Çeven et al., Phys. Rev. A 106, 063320 (2022)], this thesis investigates the application of neural-network quantum states to a two leg Bose-Hubbard ladder in the presence of strong synthetic magnetic fields. This paper showcased the reliability of variational neural networks, such as restricted Boltzmann machines and feedforward neural networks, in accurately predicting the phase diagram exhibiting superfluid-Mott insulator phase transition under strong interaction. Moreover, the neural networks successfully identified other intriguing many-body phases in the weakly interacting regime. These exciting findings firmly designate a two-leg Bose-Hubbard ladder with magnetic flux as an ideal testbed for advancing the field of neural-network quantum states. By expanding these previous results, this thesis contains various essential aspects, including a comprehensive introduction and analysis of the vanilla Bose-Hubbard model and the two-leg Bose-Hubbard ladder under magnetic flux, an in-depth overview of neural-network quantum states tailored for bosonic systems, and a thorough presentation and analysis of the obtained results using neural-network quantum states for these two Bose-Hubbard models.
Open Access
A new initialization technique: truncated towers
(IEEE, 2022-08-29) Çatalbaş, Burak; Morgül, Ömer
Artificial Neural Networks (ANN) can perform various tasks by modifying their parameters, determined by initialization methods, using activation functions and learning algorithms. Alongside chosen training dataset, suitability and success of the chosen methods determine how well they can fulfill these tasks. With the ‘Truncated Tower Distribution’, which we developed to find a better initialization method, a partial and visible increase in success rates have been achieved in classification tasks on different data sets, using neuron number-independent and neuron-number-dependent initialization methods.
Open Access
Prediction of cryptocurrency returns using machine learning
(Springer, 2021-02) Akyildirim, E.; Goncu, A.; Sensoy, Ahmet
In this study, the predictability of the most liquid twelve cryptocurrencies are analyzed at the daily and minute level frequencies using the machine learning classification algorithms including the support vector machines, logistic regression, artificial neural networks, and random forests with the past price information and technical indicators as model features. The average classification accuracy of four algorithms are consistently all above the 50% threshold for all cryptocurrencies and for all the timescales showing that there exists predictability of trends in prices to a certain degree in the cryptocurrency markets. Machine learning classification algorithms reach about 55–65% predictive accuracy on average at the daily or minute level frequencies, while the support vector machines demonstrate the best and consistent results in terms of predictive accuracy compared to the logistic regression, artificial neural networks and random forest classification algorithms.
Open Access
Recognizing targets from infrared intensity scan patterns using artificial neural networks
(S P I E - International Society for Optical Engineering, 2009-01-30) Ayta̧ç, T.; Barshan, B.
This study investigates the use of simple, low-cost infrared sensors for the recognition of geometry and surface type of commonly encountered features or targets in indoor environments, such as planes, corners, and edges. The intensity measurements obtained from such sensors are highly dependent on the location, geometry, and surface properties of the reflecting target in a way that cannot be represented by a simple analytical relationship, therefore complicating the localization and recognition process. We employ artificial neural networks to determine the geometry and the surface type of targets and provide experimental verification with three different geometries and three different surface types. The networks are trained with the Levenberg-Marquardt algorithm and pruned with the optimal brain surgeon technique. The geometry and the surface type of targets can be correctly classified with rates of 99 and 78.4%, respectively. An average correct classification rate of 78% is achieved when both geometry and surface type are differentiated. This indicates that the geometrical properties of the targets are more distinctive than their surface properties, and surface determination is the limiting factor in recognizing the patterns. The results demonstrate that processing the data from simple infrared sensors through suitable techniques can help us exploit their full potential and extend their usage beyond well-known applications.
Open Access
Time-aware and context-sensitive ensemble learning for sequential data
(Institute of Electrical and Electronics Engineers, 2023-09-26) Fazla, Arda; Aydın, Mustafa E.; Kozat, Suleyman Serdar
We investigate sequential time series data through ensemble learning. Conventional ensemble algorithms and the recently introduced ones have provided significant performance improvements in widely publicized time series prediction competitions for stationary data. However, recent studies are inadequate in capturing the temporally varying statistics for non-stationary data. To this end, we introduce a novel approach using a meta learner that effectively combines base learners in both a time varying and context-dependent manner. Our approach is based on solving a weight optimization problem that minimizes a specific loss function with constraints on the linear combination of the base learners. The constraints are theoretically analyzed under known statistics and integrated into the learning procedure of the meta-learner as part of the optimization in an automated manner. We demonstrate significant performance improvements on real-life data and well-known competition datasets over the widely used conventional ensemble methods and the state-ofthe-art forecasting methods in the machine learning literature. Furthermore, we openly share the source code of our method to facilitate further research and comparison.