Pub Date : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333991
JooYong Shim, Joongheon Kim, Jong-Kook Kim
The trade-off between accuracy and computation should be considered when applying generative adversarial network (GAN)-based image generation to real-world applications. This paper presents a simple yet efficient method based on Progressive Growing of GANs (PGGAN) to exploit the trade-off for image generation. The scheme is evaluated using the LSUN dataset.
{"title":"On the Tradeoff between Computation-Time and Learning-Accuracy in GAN-based Super-Resolution Deep Learning","authors":"JooYong Shim, Joongheon Kim, Jong-Kook Kim","doi":"10.1109/ICOIN50884.2021.9333991","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333991","url":null,"abstract":"The trade-off between accuracy and computation should be considered when applying generative adversarial network (GAN)-based image generation to real-world applications. This paper presents a simple yet efficient method based on Progressive Growing of GANs (PGGAN) to exploit the trade-off for image generation. The scheme is evaluated using the LSUN dataset.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"1 1","pages":"422-424"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"72932958","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333979
Indraneel Ghosh, Subham Kumar, Ashutosh Bhatia, D. Vishwakarma
Command-and-Control (C&C) servers use Domain Generation Algorithms (DGAs) to communicate with bots for uploading malware and coordinating attacks. Manual detection methods and sinkholing fail to work against these algorithms, which can generate thousands of domain names within a short period. This creates a need for an automated and intelligent system that can detect such malicious domains. LSTM (Long Short Term Memory) is one of the most popularly used deep learning architectures for DGA detection, but it performs poorly against Dictionary Domain Generation Algorithms. This work explores the application of various machine learning techniques to this problem, including specialized approaches such as Auxiliary Loss Optimization for Hypothesis Augmentation (ALOHA), with a particular focus on their performance against Dictionary Domain Generation Algorithms. The ALOHA-LSTM model improves the accuracy of Dictionary Domain Generation Algorithms compared to the state of the art LSTM model. Improvements were observed in the case of word-based DGAs as well. Addressing this issue is of paramount importance, as they are used extensively in carrying out Distributed Denial-of-Service (DDoS) attacks. DDoS and its variants comprise one of the most significant and damaging cyber-attacks that have been carried out in the past.
{"title":"Using Auxiliary Inputs in Deep Learning Models for Detecting DGA-based Domain Names","authors":"Indraneel Ghosh, Subham Kumar, Ashutosh Bhatia, D. Vishwakarma","doi":"10.1109/ICOIN50884.2021.9333979","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333979","url":null,"abstract":"Command-and-Control (C&C) servers use Domain Generation Algorithms (DGAs) to communicate with bots for uploading malware and coordinating attacks. Manual detection methods and sinkholing fail to work against these algorithms, which can generate thousands of domain names within a short period. This creates a need for an automated and intelligent system that can detect such malicious domains. LSTM (Long Short Term Memory) is one of the most popularly used deep learning architectures for DGA detection, but it performs poorly against Dictionary Domain Generation Algorithms. This work explores the application of various machine learning techniques to this problem, including specialized approaches such as Auxiliary Loss Optimization for Hypothesis Augmentation (ALOHA), with a particular focus on their performance against Dictionary Domain Generation Algorithms. The ALOHA-LSTM model improves the accuracy of Dictionary Domain Generation Algorithms compared to the state of the art LSTM model. Improvements were observed in the case of word-based DGAs as well. Addressing this issue is of paramount importance, as they are used extensively in carrying out Distributed Denial-of-Service (DDoS) attacks. DDoS and its variants comprise one of the most significant and damaging cyber-attacks that have been carried out in the past.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"78 1","pages":"391-396"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"77067445","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333854
Yeunwoong Kyung
Software-defined access networks (SDAN) have gained considerable attention due to the flexible and fine-granular mobile traffic management. Due to the dynamic mobility feature, an efficient flow rule management method is required in SDAN. To deal with the challenges of the mobility feature and limited rule space in forwarding nodes, a mobility-aware prioritized flow rule placement scheme in SDAN is proposed. The proposed scheme proactively performs the flow rule placement based on the flow characteristics considering delay-sensitiveness since it can directly affect users’ QoS experiences.
{"title":"Mobility-Aware Prioritized Flow Rule Placement in Software-Defined Access Networks","authors":"Yeunwoong Kyung","doi":"10.1109/ICOIN50884.2021.9333854","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333854","url":null,"abstract":"Software-defined access networks (SDAN) have gained considerable attention due to the flexible and fine-granular mobile traffic management. Due to the dynamic mobility feature, an efficient flow rule management method is required in SDAN. To deal with the challenges of the mobility feature and limited rule space in forwarding nodes, a mobility-aware prioritized flow rule placement scheme in SDAN is proposed. The proposed scheme proactively performs the flow rule placement based on the flow characteristics considering delay-sensitiveness since it can directly affect users’ QoS experiences.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"43 1","pages":"59-61"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"77790555","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333993
Jin-Oh Park, Dae-Young Lee, Young-Seok Choi
In the world, epilepsy is a common neurological disorder, and around 50 million people have epilepsy. The risk of premature death in epileptic patients is up to 3 times higher than the general population. To improve epilepsy patients’ quality of life, the use of non-invasive brain rhythm, i.e., electroencephalogram (EEG) has an important role in detecting an epileptic seizure that is the hallmark of epilepsy. By measuring the complexity of the EEG signals from patients, various entropy methods are used for detecting a variety of types of epileptic seizures. Conventional entropy methods such as the Approximate Entropy (ApEn) and Sample Entropy (SampEn) are dependent on data length and predetermined parameters. Here, we present a multiscale extension of Distribution Entropy (DistEn) that addresses the disadvantages of conventional entropy measures, which is referred to as multiscale DistEn (MDE). The proposed MDE is composed of a moving averaging procedure and DistEn estimation to reflect the reliable complexities over multiple temporal scales for short length EEG signals. The validation of the performance of MDE using actual normal and epileptic EEG signals is carried out. The experimental results show that MDE yields an outstanding performance in distinguishing the ictal EEG recordings compared to other entropy measures for short EEG recordings.
{"title":"Robust Epileptic Seizure Detection Using Multiscale Distribution Entropy Analysis for Short EEG Recordings","authors":"Jin-Oh Park, Dae-Young Lee, Young-Seok Choi","doi":"10.1109/ICOIN50884.2021.9333993","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333993","url":null,"abstract":"In the world, epilepsy is a common neurological disorder, and around 50 million people have epilepsy. The risk of premature death in epileptic patients is up to 3 times higher than the general population. To improve epilepsy patients’ quality of life, the use of non-invasive brain rhythm, i.e., electroencephalogram (EEG) has an important role in detecting an epileptic seizure that is the hallmark of epilepsy. By measuring the complexity of the EEG signals from patients, various entropy methods are used for detecting a variety of types of epileptic seizures. Conventional entropy methods such as the Approximate Entropy (ApEn) and Sample Entropy (SampEn) are dependent on data length and predetermined parameters. Here, we present a multiscale extension of Distribution Entropy (DistEn) that addresses the disadvantages of conventional entropy measures, which is referred to as multiscale DistEn (MDE). The proposed MDE is composed of a moving averaging procedure and DistEn estimation to reflect the reliable complexities over multiple temporal scales for short length EEG signals. The validation of the performance of MDE using actual normal and epileptic EEG signals is carried out. The experimental results show that MDE yields an outstanding performance in distinguishing the ictal EEG recordings compared to other entropy measures for short EEG recordings.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"2 1","pages":"473-476"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"82654372","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333970
Trong-Hop Do, Dang-Khoa Tran, Dinh-Quang Hoang, Dat Vuong, Trong-Minh Hoang, Nhu-Ngoc Dao, Chunghyun Lee, Sungrae Cho
Intelligent Transport System (ITS) has been considered is the ultimate goal of traffic management in the 21st century. ITS is hoped to create a more efficient transport system and safer traffic experience. An ITS comprises many components of which traffic data collection is one of the essential functionalities. This data collection component is responsible for collecting various kinds of data on which the system relies to make responses to traffic conditions. One of the most important data to be collected is vehicle speed. With the rapid development of artificial intelligence, computer vision based techniques have been used increasingly for vehicle speed estimation. However, most techniques focus on daytime environment. This paper proposes a novel algorithm for vehicle speed estimation. Transfer learning with YOLO is used as the backbone algorithm for detecting the vehicle taillights. Based on the distance between two taillights, a model that combines camera geometry and Kalman filters is proposed to estimate the vehicle speed. The advantage of the proposed algorithm is that it can quickly estimate the vehicle speed without prerequisite information about the vehicle which to be known as in many existing algorithms. Furthermore, the processing time of the proposed algorithm is very fast thanks to the backbone deep learning model. Owing to the Kalman filter, the proposed algorithm can achieve very high level of speed estimation accuracy. In this paper, the performance of the proposed algorithm is verified through experiment results.
{"title":"A Novel Algorithm for Estimating Fast-Moving Vehicle Speed in Intelligent Transport Systems","authors":"Trong-Hop Do, Dang-Khoa Tran, Dinh-Quang Hoang, Dat Vuong, Trong-Minh Hoang, Nhu-Ngoc Dao, Chunghyun Lee, Sungrae Cho","doi":"10.1109/ICOIN50884.2021.9333970","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333970","url":null,"abstract":"Intelligent Transport System (ITS) has been considered is the ultimate goal of traffic management in the 21st century. ITS is hoped to create a more efficient transport system and safer traffic experience. An ITS comprises many components of which traffic data collection is one of the essential functionalities. This data collection component is responsible for collecting various kinds of data on which the system relies to make responses to traffic conditions. One of the most important data to be collected is vehicle speed. With the rapid development of artificial intelligence, computer vision based techniques have been used increasingly for vehicle speed estimation. However, most techniques focus on daytime environment. This paper proposes a novel algorithm for vehicle speed estimation. Transfer learning with YOLO is used as the backbone algorithm for detecting the vehicle taillights. Based on the distance between two taillights, a model that combines camera geometry and Kalman filters is proposed to estimate the vehicle speed. The advantage of the proposed algorithm is that it can quickly estimate the vehicle speed without prerequisite information about the vehicle which to be known as in many existing algorithms. Furthermore, the processing time of the proposed algorithm is very fast thanks to the backbone deep learning model. Owing to the Kalman filter, the proposed algorithm can achieve very high level of speed estimation accuracy. In this paper, the performance of the proposed algorithm is verified through experiment results.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"24 1","pages":"499-503"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"82099961","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333853
Myungjae Shin, David A. Mohaisen, Joongheon Kim
Time series prediction plays a significant role in the Bitcoin market because of volatile characteristics. Recently, deep neural networks with advanced techniques such as ensembles have led to studies that show successful performance in various fields. In this paper, an ensemble-enabled Long Short-Term Memory (LSTM) with various time interval models is proposed for predicting Bitcoin price. Although hour and minute data set are shown to provide moderate shifts, daily data has relatively a deterministic shift. As such, the ensemble-enabled LSTM network architecture learned the individual characteristics and impact on price predictions from each data set. Experimental results with real-world measurement data show that this learning architecture effectively forecasts prices, especially in risky time such as sudden price fall.
{"title":"Bitcoin Price Forecasting via Ensemble-based LSTM Deep Learning Networks","authors":"Myungjae Shin, David A. Mohaisen, Joongheon Kim","doi":"10.1109/ICOIN50884.2021.9333853","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333853","url":null,"abstract":"Time series prediction plays a significant role in the Bitcoin market because of volatile characteristics. Recently, deep neural networks with advanced techniques such as ensembles have led to studies that show successful performance in various fields. In this paper, an ensemble-enabled Long Short-Term Memory (LSTM) with various time interval models is proposed for predicting Bitcoin price. Although hour and minute data set are shown to provide moderate shifts, daily data has relatively a deterministic shift. As such, the ensemble-enabled LSTM network architecture learned the individual characteristics and impact on price predictions from each data set. Experimental results with real-world measurement data show that this learning architecture effectively forecasts prices, especially in risky time such as sudden price fall.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"10 1","pages":"603-608"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"82434705","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333882
Sheikh Salman Hassan, Umer Majeed, C. Hong
Maritime network traffic is increasing due to the ongoing need for trade and tourism, thus increasing the demand for convenient, reliable, energy-efficient, and high-speed network access at sea that could be analogous to terrestrial networks. Therefore, to ensure the concept of a connected world under the umbrella of sixth-generation (6G) networks, we propose the next-generation integrated space-oceanic network, which consists of a set of LEO satellites and marine user equipments (MUE). This paper investigates network profit maximization (NPM) by optimizing the MUE association and its resource allocation in downlink communication. The formulated optimization problem corresponds to mixed-integer nonlinear programming (MINLP). To solve this problem, we propose an iterative algorithm based on Bender’s decomposition (BD). Numerical results are provided to demonstrate the convergence and effectiveness of our proposed algorithm.
{"title":"Reliable Integrated Space-Oceanic Network Profit Maximization by Bender Decomposition Approach","authors":"Sheikh Salman Hassan, Umer Majeed, C. Hong","doi":"10.1109/ICOIN50884.2021.9333882","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333882","url":null,"abstract":"Maritime network traffic is increasing due to the ongoing need for trade and tourism, thus increasing the demand for convenient, reliable, energy-efficient, and high-speed network access at sea that could be analogous to terrestrial networks. Therefore, to ensure the concept of a connected world under the umbrella of sixth-generation (6G) networks, we propose the next-generation integrated space-oceanic network, which consists of a set of LEO satellites and marine user equipments (MUE). This paper investigates network profit maximization (NPM) by optimizing the MUE association and its resource allocation in downlink communication. The formulated optimization problem corresponds to mixed-integer nonlinear programming (MINLP). To solve this problem, we propose an iterative algorithm based on Bender’s decomposition (BD). Numerical results are provided to demonstrate the convergence and effectiveness of our proposed algorithm.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"11 11 1","pages":"565-570"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"83794518","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333860
C. Bouras, A. Gkamas, V. Kokkinos, Nikolaos Papachristos
Internet of Things (IoT) and wireless technologies like LoRa brought more opportunities for application development in a plethora of different fields. One of these is location estimation of real-time objects and people. In this study, we focus on monitoring user’s location through a wearable IoT device with LoRa connectivity. The paper presents the development and integration of an IoT ecosystem (Hardware and Software) which can be used in Search and Rescue (SAR) use cases. The proposed IoT ecosystem is evaluated and deployed in real-scenarios with established gateways. After that we compare the existed location-estimation methods in terms of attenuation problem, cost and operation as well to conclude to the most suitable solution that can be integrated in LoRaWAN environments. Finally, the conclusions of this work and improvements for possible future activity are described.
{"title":"Real-Time Geolocation Approach through LoRa on Internet of Things","authors":"C. Bouras, A. Gkamas, V. Kokkinos, Nikolaos Papachristos","doi":"10.1109/ICOIN50884.2021.9333860","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333860","url":null,"abstract":"Internet of Things (IoT) and wireless technologies like LoRa brought more opportunities for application development in a plethora of different fields. One of these is location estimation of real-time objects and people. In this study, we focus on monitoring user’s location through a wearable IoT device with LoRa connectivity. The paper presents the development and integration of an IoT ecosystem (Hardware and Software) which can be used in Search and Rescue (SAR) use cases. The proposed IoT ecosystem is evaluated and deployed in real-scenarios with established gateways. After that we compare the existed location-estimation methods in terms of attenuation problem, cost and operation as well to conclude to the most suitable solution that can be integrated in LoRaWAN environments. Finally, the conclusions of this work and improvements for possible future activity are described.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"61 4 1","pages":"186-191"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"83974667","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9333940
Jeongmin Chae, Songnam Hong
Online multiple kernel learning (OMKL) has provided an attractive performance in nonlinear function learning tasks. Leveraging a random feature (RF) approximation, the major drawback of OMKL, known as the curse of dimensionality, has been recently alleviated. These advantages enable RF-based OMKL to be considered in practice. In this paper we introduce a new research problem, named stream-based active multiple kernel learning (AMKL), where a learner is allowed to label some selected data from an oracle according to a selection criterion. This is necessary in many real-world applications since acquiring a true label is costly or time-consuming. We theoretically prove that the proposed AMKL achieves an optimal sublinear regret $mathcal{O}(sqrt{T})$ as in OMKL with little labeled data, implying that the proposed selection criterion indeed avoids unnecessary label-requests.
{"title":"Stream-Based Active Learning with Multiple Kernels","authors":"Jeongmin Chae, Songnam Hong","doi":"10.1109/ICOIN50884.2021.9333940","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9333940","url":null,"abstract":"Online multiple kernel learning (OMKL) has provided an attractive performance in nonlinear function learning tasks. Leveraging a random feature (RF) approximation, the major drawback of OMKL, known as the curse of dimensionality, has been recently alleviated. These advantages enable RF-based OMKL to be considered in practice. In this paper we introduce a new research problem, named stream-based active multiple kernel learning (AMKL), where a learner is allowed to label some selected data from an oracle according to a selection criterion. This is necessary in many real-world applications since acquiring a true label is costly or time-consuming. We theoretically prove that the proposed AMKL achieves an optimal sublinear regret $mathcal{O}(sqrt{T})$ as in OMKL with little labeled data, implying that the proposed selection criterion indeed avoids unnecessary label-requests.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"21 1","pages":"718-722"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"87134026","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 : 2021-01-13DOI: 10.1109/ICOIN50884.2021.9334003
Y. Kim, Eunwoo Kim
Incremental deep learning aims to learn a sequence of tasks while avoiding forgetting their knowledge. One naïve approach using a deep architecture is to increase the capacity of the architecture as the number of tasks increases. However, this is followed by heavy memory consumption and makes the approach not practical. If we attempt to avoid such an issue with a fixed capacity, we encounter another challenging problem called catastrophic forgetting, which leads to a notable degradation of performance on previously learned tasks. To overcome these problems, we propose a clustering-guided incremental learning approach that can mitigate catastrophic forgetting while not increasing the capacity of an architecture. The proposed approach adopts a parameter-splitting strategy to assign a subset of parameters in an architecture for each task to prevent forgetting. It uses a clustering approach to discover the relationship between tasks by storing a few samples per task. When we learn a new task, we utilize the knowledge of the relevant tasks together with the current task to improve performance. This approach could maximize the efficiency of the approach realized in a single fixed architecture. Experimental results with a number of fine-grained datasets show that our method outperforms existing competitors.
{"title":"Clustering-Guided Incremental Learning of Tasks","authors":"Y. Kim, Eunwoo Kim","doi":"10.1109/ICOIN50884.2021.9334003","DOIUrl":"https://doi.org/10.1109/ICOIN50884.2021.9334003","url":null,"abstract":"Incremental deep learning aims to learn a sequence of tasks while avoiding forgetting their knowledge. One naïve approach using a deep architecture is to increase the capacity of the architecture as the number of tasks increases. However, this is followed by heavy memory consumption and makes the approach not practical. If we attempt to avoid such an issue with a fixed capacity, we encounter another challenging problem called catastrophic forgetting, which leads to a notable degradation of performance on previously learned tasks. To overcome these problems, we propose a clustering-guided incremental learning approach that can mitigate catastrophic forgetting while not increasing the capacity of an architecture. The proposed approach adopts a parameter-splitting strategy to assign a subset of parameters in an architecture for each task to prevent forgetting. It uses a clustering approach to discover the relationship between tasks by storing a few samples per task. When we learn a new task, we utilize the knowledge of the relevant tasks together with the current task to improve performance. This approach could maximize the efficiency of the approach realized in a single fixed architecture. Experimental results with a number of fine-grained datasets show that our method outperforms existing competitors.","PeriodicalId":6741,"journal":{"name":"2021 International Conference on Information Networking (ICOIN)","volume":"1 1","pages":"417-421"},"PeriodicalIF":0.0,"publicationDate":"2021-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"87906431","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：