• Journal of Intelligence and Information Systems
• /
• v.23 no.2
• /
• pp.1-17
• /
• 2017

The deep learning framework is software designed to help develop deep learning models. Some of its important functions include "automatic differentiation" and "utilization of GPU". The list of popular deep learning framework includes Caffe (BVLC) and Theano (University of Montreal). And recently, Microsoft's deep learning framework, Microsoft Cognitive Toolkit, was released as open-source license, following Google's Tensorflow a year earlier. The early deep learning frameworks have been developed mainly for research at universities. Beginning with the inception of Tensorflow, however, it seems that companies such as Microsoft and Facebook have started to join the competition of framework development. Given the trend, Google and other companies are expected to continue investing in the deep learning framework to bring forward the initiative in the artificial intelligence business. From this point of view, we think it is a good time to compare some of deep learning frameworks. So we compare three deep learning frameworks which can be used as a Python library. Those are Google's Tensorflow, Microsoft's CNTK, and Theano which is sort of a predecessor of the preceding two. The most common and important function of deep learning frameworks is the ability to perform automatic differentiation. Basically all the mathematical expressions of deep learning models can be represented as computational graphs, which consist of nodes and edges. Partial derivatives on each edge of a computational graph can then be obtained. With the partial derivatives, we can let software compute differentiation of any node with respect to any variable by utilizing chain rule of Calculus. First of all, the convenience of coding is in the order of CNTK, Tensorflow, and Theano. The criterion is simply based on the lengths of the codes and the learning curve and the ease of coding are not the main concern. According to the criteria, Theano was the most difficult to implement with, and CNTK and Tensorflow were somewhat easier. With Tensorflow, we need to define weight variables and biases explicitly. The reason that CNTK and Tensorflow are easier to implement with is that those frameworks provide us with more abstraction than Theano. We, however, need to mention that low-level coding is not always bad. It gives us flexibility of coding. With the low-level coding such as in Theano, we can implement and test any new deep learning models or any new search methods that we can think of. The assessment of the execution speed of each framework is that there is not meaningful difference. According to the experiment, execution speeds of Theano and Tensorflow are very similar, although the experiment was limited to a CNN model. In the case of CNTK, the experimental environment was not maintained as the same. The code written in CNTK has to be run in PC environment without GPU where codes execute as much as 50 times slower than with GPU. But we concluded that the difference of execution speed was within the range of variation caused by the different hardware setup. In this study, we compared three types of deep learning framework: Theano, Tensorflow, and CNTK. According to Wikipedia, there are 12 available deep learning frameworks. And 15 different attributes differentiate each framework. Some of the important attributes would include interface language (Python, C ++, Java, etc.) and the availability of libraries on various deep learning models such as CNN, RNN, DBN, and etc. And if a user implements a large scale deep learning model, it will also be important to support multiple GPU or multiple servers. Also, if you are learning the deep learning model, it would also be important if there are enough examples and references.

• Journal of Intelligence and Information Systems
• /
• v.18 no.3
• /
• pp.97-117
• /
• 2012

Nowadays, Internet users tend to do a variety of actions at the same time such as web browsing, social networking and multimedia consumption. While watching a video, once a user is interested in any product, the user has to do information searches to get to know more about the product. With a conventional approach, user has to search it separately with search engines like Bing or Google, which might be inconvenient and time-consuming. For this reason, a video annotation platform has been developed in order to provide users more convenient and more interactive ways with video content. In the future of smart TV environment, users can follow annotated information, for example, a link to a vendor to buy the product of interest. It is even better to enable users to search for information by directly discussing with friends. Users can effectively get useful and relevant information about the product from friends who share common interests or might have experienced it before, which is more reliable than the results from search engines. Social networking services provide an appropriate environment for people to share products so that they can show new things to their friends and to share their personal experiences on any specific product. Meanwhile, they can also absorb the most relevant information about the product that they are interested in by either comments or discussion amongst friends. However, within a very huge graph of friends, determining the most appropriate persons to ask for information about a specific product has still a limitation within the existing conventional approach. Once users want to share or discuss a product, they simply share it to all friends as new feeds. This means a newly posted article is blindly spread to all friends without considering their background interests or knowledge. In this way, the number of responses back will be huge. Users cannot easily absorb the relevant and useful responses from friends, since they are from various fields of interest and knowledge. In order to overcome this limitation, we propose a method to filter a user's friends for information search, which leverages semantic video annotation and social networking services. Our method filters and brings out who can give user useful information about a specific product. By examining the existing Facebook information regarding users and their social graph, we construct a user profile of product interest. With user's permission and authentication, user's particular activities are enriched with the domain-specific ontology such as GoodRelations and BestBuy Data sources. Besides, we assume that the object in the video is already annotated using Linked Data. Thus, the detail information of the product that user would like to ask for more information is retrieved via product URI. Our system calculates the similarities among them in order to identify the most suitable friends for seeking information about the mentioned product. The system filters a user's friends according to their score which tells the order of whom can highly likely give the user useful information about a specific product of interest. We have conducted an experiment with a group of respondents in order to verify and evaluate our system. First, the user profile accuracy evaluation is conducted to demonstrate how much our system constructed user profile of product interest represents user's interest correctly. Then, the evaluation on filtering method is made by inspecting the ranked results with human judgment. The results show that our method works effectively and efficiently in filtering. Our system fulfills user needs by supporting user to select appropriate friends for seeking useful information about a specific product that user is curious about. As a result, it helps to influence and convince user in purchase decisions.

• Journal of the Korean Society of Radiology
• /
• v.12 no.6
• /
• pp.755-762
• /
• 2018

The purpose of this study was to compare the FA(faractional anisotropy) and ADC(apparent diffusion coefficient) values, which were derived from diffusion tensor imaging in breast cancer patients. The diffusion gradient used in this study was derived from quantitative diffusion indices using 20 directions(b-value, 0 and $1,000s/mm^2$). Quantitative analysis was analyzed using Pearson's correction and qualitative analysis using for correction coefficients. As a result, $FA_{min}$, $FA_{mean}$ and $FA_{max}$ were $0.098{\pm}0.065$, $0.302{\pm}0.142$ and $0.634{\pm}0.236$, respectively(p > 0.05). The $ADC_{min}$, $ADC_{mean}$ and $ADC_{max}$ were $0.741{\pm}0.403$, $1.095{\pm}0.394$ and $1.530{\pm}0.447$, respectively(p > 0.05). The $FA_{min}$, $FA_{mean}$, and $FA_{max}$ mean values were $0.132{\pm}0.050$, $0.418{\pm}0.094$, and $0.770{\pm}0.164$ for Category 6 and Kinetic Curve Pattern III, respectively. $ADC_{min}$, $ADC_{mean}$, and $ADC_{max}$ were $0.753{\pm}0.189$, $1.120{\pm}0.236$, and $1.615{\pm}0.372$, respectively. Quantitative analysis showed negative correlation between $ADC_{mean}-FA_{mean}$ and $ADC_{max}-FA_{max}$(p = 0.001, 0.003). The qalitative analysis showed ADC 0.628(p = 0.001), FA 0.620(p = 0.001) in the internal evaluations, ADC 0.677(p = 0.001), FA 0.695(p = 0.001) in external evaluations. In conclusion, based on the morphological examination, time to signal intensity graph is the form of wash-out(pattern III) in the dynamic contrast enhance examination, As a result, the $ADC_{mean}$ $1.120{\pm}0.236$ and $FA_{mean}$ values were $0.032{\pm}0.142$ with a negative correlation (Y=1.44-1.12X). Therefore, If we understand the shape of time to signal intensity graph and the relationship between ADC and FA, It will be a criterion for distinguishing malignant diseases in breast cancer.

• Journal of Life Science
• /
• v.26 no.8
• /
• pp.904-910
• /
• 2016

Complex traits are determined by the combined effects of many loci and are affected by gene networks or biological pathways. Systems biology approaches have an important role in the identification of candidate genes related to complex diseases or traits at the system level. The gene network analysis has been performed by diverse types of methods such as gene co-expression, gene regulatory relationships, protein-protein interaction (PPI) and genetic networks. Moreover, the network-based methods were described for predicting gene functions such as graph theoretic method, neighborhood counting based methods and weighted function. However, there are a limited number of researches in livestock. The present study systemically analyzed genes associated with 102 types of economic traits based on the Animal Trait Ontology (ATO) and identified their relationships based on the gene co-expression network and PPI network in cattle. Then, we constructed the two types of gene network databases and network visualization system (http://www.nabc.go.kr/cg). We used a gene co-expression network analysis from the bovine expression value of bovine genes to generate gene co-expression network. PPI network was constructed from Human protein reference database based on the orthologous relationship between human and cattle. Finally, candidate genes and their network relationships were identified in each trait. They were typologically centered with large degree and betweenness centrality (BC) value in the gene network. The ontle program was applied to generate the database and to visualize the gene network results. This information would serve as valuable resources for exploiting genomic functions that influence economically and agriculturally important traits in cattle.

• Asia pacific journal of information systems
• /
• v.18 no.3
• /
• pp.67-90
• /
• 2008

Prior to making choices among online products and services, consumers often search online product reviews written by other consumers. Online product reviews have great influences on consumer behavior because they are believed to be more reliable than information provided by sellers. However, ever-increasing lists of product reviews make it difficult for consumers to find the right information efficiently. A customized search mechanism is a method to provide personalized information which fits the user's requirements. This study examines effects of a customized search mechanism and perceived similarity between consumers and product reviewers on consumer behaviors. More specifically, we address the following research questions: (1) Can a customized search mechanism increase perceived similarity between product review authors and readers? (2) Are product reviews perceived as more credible when product reviews were written by the authors perceived similar to them? (3) Does credibility of product reviews have a positive impact on acceptance of product reviews? (4) Does acceptance of product reviews have an influence on purchase intention of the readers? To examine these research questions, a lab experiment with a between-subject factor (whether a customized search mechanism is provided or not) design was employed. In order to enhance mundane realism and increase generalizability of the findings, the experiment sites were built based on a real online store, cherrya.com (http://www.cherrya.com/). Sixty participants were drawn from a pool that consisted of undergraduate and graduate students in a large university. Participation was voluntary; all the participants received 5,000 won to encourage their motivation and involvement in the experiment tasks. In addition, 15 participants, who selected by a random draw, received 30,000 won to actually purchase the product that he or she decided to buy during the experiment. Of the 60 participants, 25 were male and 35 were female. In examining the homogeneity between the two groups, the results of t-tests revealed no significant difference in gender, age, academic years, online shopping experience, and Internet usage. To test our research model, we completed tests of the measurement models and the structural models using PLS Graph version 3.00. The analysis confirmed individual item reliability, internal consistency, and discriminant validity of measurements. The results show that participants feel more credible when product reviews were written by the authors perceived similar to them, credibility of product reviews have a positive impact on acceptance of product reviews, and acceptance of product reviews have an influence on purchase intention of the readers. However, a customized search mechanism did not increase perceived similarity between product review authors and readers. The results imply that there is an urgent need to develop a better customized search tool in order to increase perceived similarity between product review authors and readers.

• Journal of Korean Society of Environmental Engineers
• /
• v.38 no.2
• /
• pp.47-55
• /
• 2016

In this study, WUA (Weighted Usable Area) based on the Instream Flow Incremental Methodology (IFIM) was calculated to determine ecological flow at JeonJu-Cheon by using River2D model. To calibrate River2D, simulation results for low flow conditions of River2D were compared with calibrated HEC-RAS simulation results and the optimum parameters were determined. The results were RMSE (0.18), NSE (0.71) and coefficient of determination (0.78) for velocity and RMSE (0.02), NSE (0.71), coefficient of determination (0.73) for water depth. The result shows that the model successfully simulates the water flows. A selected target fish species to build the habitat suitability index were composed of Zaccoplatypus and Coreoleuciscus splendidus. These species showed the highest occurrences over the past decade in f ish monitoring. Also, The WUA-Discharge curve was calculated with the suitability index in a medium flow conditions. From the result, WUA is changed according to flowrate. In the flowrate-WUA/A graph, ecological flow can be determined at $1.8{\sim}2.0m^3/s$ for Zaccoplatypus $2.0m^3/s$ and Coreoleuciscus splendidus $1.8m^3/s$ at JeonJu-Cheon upstream. When compared with flow-duration analysis, it is demonstrative that simulation results fitted ecological flow considering quantity of available habitat for each fish species.

• KSCE Journal of Civil and Environmental Engineering Research
• /
• v.32 no.3B
• /
• pp.153-159
• /
• 2012

In this study, we proposed modified settling basins transformed by substituting the downstream sill for low head weirs such as generic labyrinth weir and inclined crest labyrinth weir worked as internal baffles. Laboratory experiments were carried out to understand hydraulic characteristics inside of the settling basin to improve the efficiency of sedimentation rate. For a quantitative analysis, we suggested the headwater ratio($H_t/P$), the magnification ratio(L/W) and the inflow rate per total crest length($q_L$) as primary analysis indexes for sedimentation efficiency. Six different types of settling basin were used for labscaled pilot tests by distinguishing with internal structures. Based on results, the variation of headwater ratio with the change of magnification ratio would highly affect the deposition efficiency(%) and it was improved under specific condition that repeating arrange number(N) of labyrinth weir was between 2 and 4. Also, the regression analysis showed that initial condition and shape for improving sedimentation efficiency were plotted on the graph for both $q_L{\geq}3.5cm^2/s$ and $L/W{\leq}3.5$. It would be expected that the geometrically optimized labyrinth settling basin could be designed with proper deposition efficiency for inflow rates of influent and required area of settling basin utilizing the proposed analysis index in this study.

• Journal of Digital Convergence
• /
• v.10 no.10
• /
• pp.529-538
• /
• 2012

This study analyzed a variety of different kinds of usage patterns of four-digit personal identification numbers(PINs) based on the data collected from students attending at A-university in 2011. According to the analysis, the 4 digit PINs '1234' was being used most frequently which is similar to the findings of the study by the Daniel Amitay research. In addition, almost 50 percent of the students were using the four-digit PINs closely related to their personal information, and more than 10 percent of them were using it only by the reason of easiness to remember or convenience to use. Number '0' was the most frequently used, and girl students used it more than boys did. According to the result of analysis of four-digit PINs, it was found that number '1' was the most frequently used in the first and the third position. It was also found that number '2' and number '4' were the most frequently used in the second and the forth position respectively. Among students who had used number 3 in third position, around 63.6 percent of those students were using number 4 in the forth position in their PINs which showed highest frequency from analysis of double-digits connected together in each position. The analysis of PINs were represented by the polygonal (type) graph with that X-axis showed from first to forth position and Y-axis showed number had been used at each position. Among many polygonal graphs, the form showed an N-type which was called in type 1 took the first place by 35 percent.

• Journal of the Korean Society for Railway
• /
• v.19 no.3
• /
• pp.331-340
• /
• 2016

Wheel load, design velocity, traffic amount (MGT), stiffness and thickness of sub-layers of asphalt concrete track are selected as performance design parameters in this study. A pseudo-static wheel load computed considering the dynamic amplification factor (DAF) based on the design velocity of the KTX was applied to the top of asphalt concrete track for full three dimensional structural analysis using the ABAQUS program. Tensile strains at the bottom of the asphalt concrete layer and vertical strains at the top of the subgrade were computed from the structural FEA with different combinations of performance parameter values for one asphalt concrete track section. Utilizing the computed structural analysis results such as the tensile strains and the vertical strains, it was possible to develop design graphs to investigate proper track sections for different combination of the performance parameters including wheel load, design velocity, traffic amount(MGT), stiffness and thickness of asphalt concrete layers for any given design life. By analyzing the proposed design graphs for asphalt concrete track, it was possible to propose simple design tables that can be used by engineers for the effective and fast design of track.

• Journal of KIISE:Software and Applications
• /
• v.31 no.3
• /
• pp.293-307
• /
• 2004

We consider the problem of scheduling tasks of a precedence constrained task graph, where each task has its execution time and deadline, onto a set of identical processors in a way that simultaneously minimizes the number of processors required and the total tardiness of tasks. Most existing approaches tend to focus on the minimization of the total tardiness of tasks. In another methods, solutions to this problem are usually computed by combining the two objectives into a simple criterion to be optimized. In this paper, the minimization is carried out using a multiobjective genetic algorithm (GA) that independently considers both criteria by using a vector-valued cost function. We present various GA components that are well suited to the problem of task scheduling, such as a non-trivial encoding strategy. a domination-based selection operator, and a heuristic crossover operator We also provide three local improvement heuristics that facilitate the fast convergence of GA's. The experimental results showed that when compared to five methods used previously, such as list-scheduling algorithms and a specific genetic algorithm, the Performance of our algorithm was comparable or better for 178 out of 180 randomly generated task graphs.