• International journal of advanced smart convergence
• /
• v.8 no.1
• /
• pp.141-146
• /
• 2019

These days, there are many applications using neural networks as parts of their system. On the other hand, adversarial examples have become an important issue concerining the security of neural networks. A classifier in neural networks can be fooled and make it miss-classified by adversarial examples. There are many research to encounter adversarial examples by using denoising methods. Some of them using GAN (Generative Adversarial Network) in order to remove adversarial noise from input images. By producing an image from generator network that is close enough to the original clean image, the adversarial examples effects can be reduced. However, there is a chance when adversarial noise can survive the approximation process because it is not like a normal noise. In this chance, we propose a research that utilizes high-level representation in the classifier by combining GAN network with a trained U-Net network. This approach focuses on minimizing the loss function on high representation terms, in order to minimize the difference between the high representation level of the clean data and the approximated output of the noisy data in the training dataset. Furthermore, the generated output is checked whether it shows minimum error compared to true label or not. U-Net network is trained with true label to make sure the generated output gives minimum error in the end. At last, the remaining adversarial noise that still exist after low-level approximation can be removed with the U-Net, because of the minimization on high representation terms.

• Journal of the Korean Society of Visualization
• /
• v.21 no.1
• /
• pp.18-25
• /
• 2023

In these days, high dimensional data prediction technology based on neural network shows compelling results in many different kind of field including engineering. Especially, a lot of variants of convolution neural network are widely utilized to develop pixel level prediction model for high dimensional data such as picture, or physical field value from the sensors. In this study, velocity vector field of ideal flow on ship surface is estimated on pixel level by Unet. First, potential flow analysis was conducted for the set of hull form data which are generated by hull form transformation method. Thereafter, four different neural network with a U-shape structure were conFig.d to train velocity vectors at the node position of pre-processed hull form data. As a result, for the test hull forms, it was confirmed that the network with short skip-connection gives the most accurate prediction results of streamlines and velocity magnitude. And the results also have a good agreement with potential flow analysis results. However, in some cases which don't have nothing in common with training data in terms of speed or shape, the network has relatively high error at the region of large curvature.

• Science of Emotion and Sensibility
• /
• v.20 no.2
• /
• pp.161-170
• /
• 2017

Emotion affects many parts of human life such as learning ability, behavior and judgment. It is important to understand human nature. Emotion can only be inferred from facial expressions or gestures, what it actually is. In particular, emotion is difficult to classify not only because individuals feel differently about emotion but also because visually induced emotion does not sustain during whole testing period. To solve the problem, we acquired bio-signals and extracted features from those signals, which offer objective information about emotion stimulus. The emotion pattern classifier was composed of unsupervised learning algorithm with hidden nodes and feature vectors. Restricted Boltzmann machine (RBM) based on probability estimation was used in the unsupervised learning and maps emotion features to transformed dimensions. The emotion was characterized by non-linear classifiers with hidden nodes of a multi layer neural network, named deep belief network (DBN). The accuracy of DBN (about 94 %) was better than that of back-propagation neural network (about 40 %). The DBN showed good performance as the emotion pattern classifier.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.18 no.3
• /
• pp.106-118
• /
• 2019

In recent years, research and development on image object recognition using artificial intelligence have been actively carried out, and it is expected to be used for road maintenance. Among them, artificial intelligence models for object detection of road surface are continuously introduced. In order to develop such object recognition algorithms, a backbone network that extracts feature maps is essential. In this paper, we will discuss how to select the appropriate neural network. To accomplish it, we compared with 4 different deep neural networks using 6,000 road surface damage images. Based on three evaluation methods for analyzing characteristics of neural networks, we propose a method to determine optimal neural networks. In addition, we improved the performance through optimal tuning of hyper-parameters, and finally developed a light backbone network that can achieve 85.9% accuracy of road surface damage classification.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.67 no.8
• /
• pp.1071-1079
• /
• 2018

In this study, the design methodology as well as network architecture of Support Vector Machine based Polynomial Neural Network, which is a kind of the dynamically generated neural networks, is introduced. The Support Vector Machine based polynomial neural networks is given as a novel network architecture redesigned with the aid of polynomial neural networks and Support Vector Machine. The generic polynomial neural networks, whose nodes are made of polynomials, are dynamically generated in each layer-wise. The individual nodes of the support vector machine based polynomial neural networks is constructed as a support vector machine, and the nodes as well as layers of the support vector machine based polynomial neural networks are dynamically generated as like the generation process of the generic polynomial neural networks. Support vector machine is well known as a sort of robust pattern classifiers. In addition, in order to enhance the structural flexibility as well as the classification performance of the proposed classifier, multi-objective particle swarm optimization is used. In other words, the optimization algorithm leads to sequentially successive generation of each layer of support vector based polynomial neural networks. The bench mark data sets are used to demonstrate the pattern classification performance of the proposed classifiers through the comparison of the generalization ability of the proposed classifier with some already studied classifiers.

• Journal of the Korean Institute of Telematics and Electronics C
• /
• v.36C no.4
• /
• pp.78-86
• /
• 1999

In this paper, a distributed arithmetic digital neural network with learning and testing phase implemented in a body has been studied. The proposed technique is based on the two facts; one is that the weighting coefficients adjusted will be stored in registers without shift, because input values or input patterns are not changed while learning and the other is that the input patterns stored in registers are not changed while testing. The proposed digital neural network is simulated by hardware description language such as VHDL and verified the performance that the neural network was applied to the recognition of seven-segment. To verify proposed neural networks, we compared the learning process of modified perceptron learning algorithm simulated by software with VHDL for 7-segment number recognizer. The results are as follows: There was a little difference in learning time and iteration numbers according to the input pattern, but generally the iteration numbers are 1000 to 10000 and the learning time is 4 to 200$\mu\textrm{s}$. So we knew that the operation of the neural network is learned in the same way with the learning of software simulation, and the proposed neural networks are properly operated. And also the implemented neural network can be built with less amounts of components compared with board system neural network.

• Proceedings of the KIEE Conference
• /
• 1990.11a
• /
• pp.450-456
• /
• 1990

In this paper, production system is implemented with the inferential neural network model using semantic network and directed graph. Production system can be implemented with the transform of knowledge representation in production system into semantic network and of semantic network into directed graph, because directed graphs can be expressed by neural matrices. A concept node should be defined by the state vector to calculated the concepts expressed by matrices. The expressional ability of neunal network depends on how the state vector is defined. In this study, state vector is overlapped and each overlapping part acts as a inheritant of concept.

• 제어로봇시스템학회:학술대회논문집
• /
• 1994.10a
• /
• pp.313-318
• /
• 1994

It is well known that the neural network can be used as an universal approximater for functions and functionals. But these theoretical results are just an existence theorem and do not lead to decide the suitable network structure. This doubfulness whether a certain network can approximate a given function or not, brings about serious stability problems when it is used to identify a system. To overcome the stability problem, We suggest successive identification and control scheme with supervisory controller which always assures the identification process within a basin of attraction of one stable equilibrium point regardless of fittness of the network.

• Journal of Korean Society of Forest Science
• /
• v.90 no.4
• /
• pp.437-444
• /
• 2001

This paper discusses applications of neural network to forest stand field data processing and determining suitable tree species for site-specific stand characteristics. For site-specific species selection, considered were 5 major coniferous species : P. densiflora for. erecta, L. leptolepis, P. koraiensis, P. densiflora, P. thunbergii. Among 1,320 sample plot data sets, 200 data sets with the highest site index (40 data sets for each species) were chosen as the test sets for investigation. Each data set includes 13 factors describing the site characteristics of the corresponding sample plot. The results of this investigation indicate high performance of neural network in data processing procedures for extracting data sets or measurement parameters without any recognizable pattern. These data sets or measurement parameters are those which have rare effect on site-specific species suitability or disturb pattern classification procedures of neural network because of unrecognizable patterns involved. Also the results have shown high potential of neural network in determining the best-suitable tree species for site characteristics. The % accuracy of the neural network model in determining the best-suitable tree species for site characteristics ranges from 77.6% to 91.8% associated with the combination of Site factors.

• Journal of the Korean Geotechnical Society
• /
• v.21 no.8
• /
• pp.73-84
• /
• 2005

A three layered neural network model was developed using back propagation algorithm to estimate the UU undrained shear strength of Korean soft soil based on the database of actual undrained shear strengths and piezocone measurements compiled from 8 sites over the Korea. The developed model was validated by comparing model predictions with measured values about new piezocone data, which were not previously employed during development of model. Performance of the neural network model was also compared with conventional empirical methods. It was found that the number of neuron in hidden layer is different for the different combination of transfer functions of neural network models. However, all piezocone neural network models are successful in inferring a complex relationship between piezocone measurements and the undrained shear strength of Korean soft soils, which give relatively high coefficients of determination ranging from 0.69 to 0.72. Since neural network model has been generalized by self-learning from database of piezocone measurements and undrained shear strength over the various sites, the developed neural network models give more precise and generally reliable undrained shear strengths than empirical approaches which still need site specific calibration.