Despite the convolutional neural network has achieved tremendous monumental success on a variety of computer vision-related tasks, it is still extremely challenging to build a neural network with doubtless reliability. Previous works have demonstrated that the deep neural network can be efficiently fooled by human imperceptible perturbation to the input, which actually revealed the instability for interpolation. Like human-beings, an ideally trained neural network should be constrained within desired inference space and maintain correctness for both interpolation and extrapolation. In this paper, we develop a technique to verify the correctness when convolutional neural networks extrapolate beyond training data distribution by generating legitimated feature broken images, and we show that the decision boundary for convolutional neural network is not well formulated based on image features for extrapolating.
{"title":"Towards Understanding The Space of Unrobust Features of Neural Networks","authors":"Bingli Liao, Takahiro Kanzaki, Danilo Vasconcellos Vargas","doi":"10.1109/CYBCONF51991.2021.9464137","DOIUrl":"https://doi.org/10.1109/CYBCONF51991.2021.9464137","url":null,"abstract":"Despite the convolutional neural network has achieved tremendous monumental success on a variety of computer vision-related tasks, it is still extremely challenging to build a neural network with doubtless reliability. Previous works have demonstrated that the deep neural network can be efficiently fooled by human imperceptible perturbation to the input, which actually revealed the instability for interpolation. Like human-beings, an ideally trained neural network should be constrained within desired inference space and maintain correctness for both interpolation and extrapolation. In this paper, we develop a technique to verify the correctness when convolutional neural networks extrapolate beyond training data distribution by generating legitimated feature broken images, and we show that the decision boundary for convolutional neural network is not well formulated based on image features for extrapolating.","PeriodicalId":231194,"journal":{"name":"2021 5th IEEE International Conference on Cybernetics (CYBCONF)","volume":"104 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-06-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115690067","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-06-08DOI: 10.1109/CYBCONF51991.2021.9464140
M. Fukumoto, Gan Haoran, Y. Hanada
Obtaining media content suited to user’s feelings is one of the essential targets of engineering. However, it is still difficult because feelings are different between users and are hard to be shown as a certain equation. As a method of a beverage, this study proposed Interactive Tabu Search (ITS) that blends source juices for creating new beverages suited to each user’s feelings. Tabu Search is one of stochastic local searches, and its properties are a continuous neighborhood search and a tabu list prohibiting cycling. A target of optimization was the ratio of the source juices. A concrete system based on the proposed ITS was constructed with the computer, Arduino, and peristaltic pumps. A tasting experiment composed of two steps was conducted. The target was delicious blended beverage. As a result, continuous increases in the fitness values related to deliciousness were observed, and a significant increase was observed in the maximum fitness. In the progress of the ratios, both different and common trends between the subjects were observed.
{"title":"A Proposal of Interactive Tabu Search for Creating Beverage by Blending Source Juices","authors":"M. Fukumoto, Gan Haoran, Y. Hanada","doi":"10.1109/CYBCONF51991.2021.9464140","DOIUrl":"https://doi.org/10.1109/CYBCONF51991.2021.9464140","url":null,"abstract":"Obtaining media content suited to user’s feelings is one of the essential targets of engineering. However, it is still difficult because feelings are different between users and are hard to be shown as a certain equation. As a method of a beverage, this study proposed Interactive Tabu Search (ITS) that blends source juices for creating new beverages suited to each user’s feelings. Tabu Search is one of stochastic local searches, and its properties are a continuous neighborhood search and a tabu list prohibiting cycling. A target of optimization was the ratio of the source juices. A concrete system based on the proposed ITS was constructed with the computer, Arduino, and peristaltic pumps. A tasting experiment composed of two steps was conducted. The target was delicious blended beverage. As a result, continuous increases in the fitness values related to deliciousness were observed, and a significant increase was observed in the maximum fitness. In the progress of the ratios, both different and common trends between the subjects were observed.","PeriodicalId":231194,"journal":{"name":"2021 5th IEEE International Conference on Cybernetics (CYBCONF)","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-06-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123575214","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-06-08DOI: 10.1109/CYBCONF51991.2021.9464153
Yuzi Jiang, Danilo Vasconcellos Vargas
Subpopulation algorithm based on novelty (SAN) has been investigated for some time and proved that it can be used for multi-objective optimization problems. It outperforms subpopulation algorithm based on general differential evolution (SAGDE) under the same framework, which highlights its special intrinsic mechanism. This intrinsic mechanism has something in common with some state-of-the-art multi-objective optimization algorithms. However, SAN has not yet proved its ability to be better than these algorithms and has not proven its ability to optimize problems with more than 5 objectives. In this paper, the advantage of SAN over other subpopulation algorithms, i.e., novelty search, is presented in detail. The similarities and differences between the intrinsic mechanisms of SAN, nondominated sorting genetic algorithm series (NSGAs) and multi-objective evolutionary algorithm based on decomposition (MOEA/D) are also analyzed. Finally, these three algorithms are evaluated on several well-known benchmark problems with more than two objectives. The result shows SAN surpassed NSGA-III (latest version in NSGAs) in 20 out of the 32 problems, surpassed MOEA/D in 26 problems in 10 runs, which preliminary proved it surpasses the State-of-the-Art.
{"title":"Preliminary Results for Subpopulation Algorithm Based on Novelty (SAN) Compared with the State of the Art","authors":"Yuzi Jiang, Danilo Vasconcellos Vargas","doi":"10.1109/CYBCONF51991.2021.9464153","DOIUrl":"https://doi.org/10.1109/CYBCONF51991.2021.9464153","url":null,"abstract":"Subpopulation algorithm based on novelty (SAN) has been investigated for some time and proved that it can be used for multi-objective optimization problems. It outperforms subpopulation algorithm based on general differential evolution (SAGDE) under the same framework, which highlights its special intrinsic mechanism. This intrinsic mechanism has something in common with some state-of-the-art multi-objective optimization algorithms. However, SAN has not yet proved its ability to be better than these algorithms and has not proven its ability to optimize problems with more than 5 objectives. In this paper, the advantage of SAN over other subpopulation algorithms, i.e., novelty search, is presented in detail. The similarities and differences between the intrinsic mechanisms of SAN, nondominated sorting genetic algorithm series (NSGAs) and multi-objective evolutionary algorithm based on decomposition (MOEA/D) are also analyzed. Finally, these three algorithms are evaluated on several well-known benchmark problems with more than two objectives. The result shows SAN surpassed NSGA-III (latest version in NSGAs) in 20 out of the 32 problems, surpassed MOEA/D in 26 problems in 10 runs, which preliminary proved it surpasses the State-of-the-Art.","PeriodicalId":231194,"journal":{"name":"2021 5th IEEE International Conference on Cybernetics (CYBCONF)","volume":"44 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-06-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131228890","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：