Pub Date : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420616
M. A. Rashid, Hossein Amirkhani
Powerful yet simple augmentation techniques have significantly helped modern deep learning-based text classifiers to become more robust in recent years. Although these augmentation methods have proven to be effective, they often utilize random or non-contextualized operations to generate new data. In this work, we modify a specific augmentation method called Easy Data Augmentation or EDA with more sophisticated text editing operations powered by masked language models such as BERT and RoBERTa to analyze the benefits or setbacks of creating more linguistically meaningful and hopefully higher quality augmentations. Our analysis demonstrates that using a masked language model for word insertion almost always achieves better results than the initial method but it comes at a cost of more time and resources which can be comparatively remedied by deploying a lighter and smaller language model like DistilBERT.
{"title":"The Effect of Using Masked Language Models in Random Textual Data Augmentation","authors":"M. A. Rashid, Hossein Amirkhani","doi":"10.1109/CSICC52343.2021.9420616","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420616","url":null,"abstract":"Powerful yet simple augmentation techniques have significantly helped modern deep learning-based text classifiers to become more robust in recent years. Although these augmentation methods have proven to be effective, they often utilize random or non-contextualized operations to generate new data. In this work, we modify a specific augmentation method called Easy Data Augmentation or EDA with more sophisticated text editing operations powered by masked language models such as BERT and RoBERTa to analyze the benefits or setbacks of creating more linguistically meaningful and hopefully higher quality augmentations. Our analysis demonstrates that using a masked language model for word insertion almost always achieves better results than the initial method but it comes at a cost of more time and resources which can be comparatively remedied by deploying a lighter and smaller language model like DistilBERT.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129375294","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420602
Saeedeh Davoudi, S. Mirzaei
Natural Language Processing (NLP) is one of the promising fields of artificial intelligence. In recent decades, high volume of text data has been generated through the Internet. This kind of data is a valuable source of information which can be used in various fields such as information retrieval, search engines, recommender systems, etc. One practical task of text mining is document classification. In this paper, we mainly focus on Persian document classification. We introduce a new feature extraction approach derived from the combination of K-means clustering and Word2Vec to acquire semantically relevant and discriminant word representations. We call our proposed approach CC-Word2Vec (Categorical Clustering-Word2Vec) since we retrain the Word2Vec model using the word clusters of each category obtained by K-Means algorithm. We use 200 documents of 5 most frequent categories of Hamshahri news dataset to evaluate our method. We pass the extracted word vectors to Multi-Layer Perceptron (MLP) and Gradient Boosting (GB) classifiers to compare the performance of the proposed approach with Term Frequency Inverse Document Frequency (TF-IDF) and Word2Vec methods. Our new approach resulted in an improvement in the obtained accuracy of Gradient Boosting and Multi-Layer Perceptron models in comparison with TF-IDF and Word2Vec techniques.
{"title":"A Semantic-based Feature Extraction Method Using Categorical Clustering for Persian Document Classification","authors":"Saeedeh Davoudi, S. Mirzaei","doi":"10.1109/CSICC52343.2021.9420602","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420602","url":null,"abstract":"Natural Language Processing (NLP) is one of the promising fields of artificial intelligence. In recent decades, high volume of text data has been generated through the Internet. This kind of data is a valuable source of information which can be used in various fields such as information retrieval, search engines, recommender systems, etc. One practical task of text mining is document classification. In this paper, we mainly focus on Persian document classification. We introduce a new feature extraction approach derived from the combination of K-means clustering and Word2Vec to acquire semantically relevant and discriminant word representations. We call our proposed approach CC-Word2Vec (Categorical Clustering-Word2Vec) since we retrain the Word2Vec model using the word clusters of each category obtained by K-Means algorithm. We use 200 documents of 5 most frequent categories of Hamshahri news dataset to evaluate our method. We pass the extracted word vectors to Multi-Layer Perceptron (MLP) and Gradient Boosting (GB) classifiers to compare the performance of the proposed approach with Term Frequency Inverse Document Frequency (TF-IDF) and Word2Vec methods. Our new approach resulted in an improvement in the obtained accuracy of Gradient Boosting and Multi-Layer Perceptron models in comparison with TF-IDF and Word2Vec techniques.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"87 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122046548","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420590
Saeid Ghasemshirazi, Ghazaleh Shirvani
Online Social Networks (OSNs) attracted millions of users in the world. OSNs made adversaries more passionate to create malware variants to subvert the cyber defence of OSNs. Through various threat vectors, adversaries persuasively lure OSN users into installing malware on their devices at an enormous scale. One of the most horrendous forms of named malware is OSNs' botnets that conceal C&C information using OSNs' accounts of unaware users. In this paper, we present GITC (Ghost In The Cloud), which uses Telegram as a C&C server to communicate with threat actors and access targets' information in an undetectable way. Furthermore, we present our implementation of GITC. We show how GITC uses the encrypted telegram Application Programming Interface (API) to cover up records of the adversary connections to the target, and we discuss why current intrusion detection systems cannot detect GITC. In the end, we run some sets of experiments that confirm the feasibility of GITC.
在线社交网络(Online Social Networks，简称osn)在全球吸引了数以百万计的用户。osn使得对手更有激情地创建恶意软件变体来破坏osn的网络防御。通过各种威胁载体，攻击者有说服力地引诱OSN用户在其设备上大规模安装恶意软件。命名恶意软件最可怕的形式之一是osn的僵尸网络，它利用不知情用户的osn帐户隐藏C&C信息。在本文中，我们提出了GITC (Ghost In The Cloud)，它使用Telegram作为C&C服务器与威胁参与者进行通信，并以不可检测的方式访问目标的信息。此外，我们还介绍了GITC的实现。我们展示了GITC如何使用加密的电报应用程序编程接口(API)来掩盖对手与目标的连接记录，并讨论了为什么当前的入侵检测系统不能检测到GITC。最后，我们进行了几组实验，验证了GITC的可行性。
{"title":"GITCBot: A Novel Approach for the Next Generation of C&C Malware","authors":"Saeid Ghasemshirazi, Ghazaleh Shirvani","doi":"10.1109/CSICC52343.2021.9420590","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420590","url":null,"abstract":"Online Social Networks (OSNs) attracted millions of users in the world. OSNs made adversaries more passionate to create malware variants to subvert the cyber defence of OSNs. Through various threat vectors, adversaries persuasively lure OSN users into installing malware on their devices at an enormous scale. One of the most horrendous forms of named malware is OSNs' botnets that conceal C&C information using OSNs' accounts of unaware users. In this paper, we present GITC (Ghost In The Cloud), which uses Telegram as a C&C server to communicate with threat actors and access targets' information in an undetectable way. Furthermore, we present our implementation of GITC. We show how GITC uses the encrypted telegram Application Programming Interface (API) to cover up records of the adversary connections to the target, and we discuss why current intrusion detection systems cannot detect GITC. In the end, we run some sets of experiments that confirm the feasibility of GITC.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"18 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124785035","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420613
M. Asadi, M. Sadeghi, A. Y. Bafghi
In rock masses, presence of cracks greatly affects the behavior of it. Obtaining the cracks is very important in specialized analysis of rock mechanics. In computer vision applications, crack segmentation task in an intricate texture such as rock mass, is difficult. Crack segmentation problem can consider as an edge detection task so we can use edge detection methods to achieve it. In this paper, we propose a multi-classifier system based on deep convolutional neural network (CNN) to predict pixel-wise cracks in rock mass images. We provide a dataset consists of 489 RGB rock mass images with manual ground truths. For training classifiers, we create two sub-datasets obtained by mentioned dataset. Also we introduce a new approach of image labeling to improve general methods. Based on the results, our method achieves F-score of 84.0, which has a best performance compared to different methods.
{"title":"A Multi-Classifier System for Rock Mass Crack Segmentation Based on Convolutional Neural Networks","authors":"M. Asadi, M. Sadeghi, A. Y. Bafghi","doi":"10.1109/CSICC52343.2021.9420613","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420613","url":null,"abstract":"In rock masses, presence of cracks greatly affects the behavior of it. Obtaining the cracks is very important in specialized analysis of rock mechanics. In computer vision applications, crack segmentation task in an intricate texture such as rock mass, is difficult. Crack segmentation problem can consider as an edge detection task so we can use edge detection methods to achieve it. In this paper, we propose a multi-classifier system based on deep convolutional neural network (CNN) to predict pixel-wise cracks in rock mass images. We provide a dataset consists of 489 RGB rock mass images with manual ground truths. For training classifiers, we create two sub-datasets obtained by mentioned dataset. Also we introduce a new approach of image labeling to improve general methods. Based on the results, our method achieves F-score of 84.0, which has a best performance compared to different methods.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"49 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129801935","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420627
E. J. Zaferani, M. Teshnehlab, Mansoor Vali
In this research, a study of automatic personality perception based on the Big-five Inventory (BFI) is done. To extract and select appropriate features for the classification, we employ an auto-encoder as a nonlinear feature learning technique. Since an auto-encoder does not extract proper classification lonely, a saddle point is found by a stop criterion based on maximum separate ability in binary classes. The results reveal that nonlinear features enhance the classification results in most personality traits. Furthermore, we use an adaptive neuro-fuzzy inference system classification to model the uncertainty rooted in mental states and affect the classification results through the extracted features. The classification outcomes on SSPNet Speaker Personality dataset demonstrate significant improvement in the results of four traits. These outgrowths verify the existence of uncertainty in the speech signal.
{"title":"Automatic Personality Perception Using Autoencoder And Hierarchical Fuzzy Classification","authors":"E. J. Zaferani, M. Teshnehlab, Mansoor Vali","doi":"10.1109/CSICC52343.2021.9420627","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420627","url":null,"abstract":"In this research, a study of automatic personality perception based on the Big-five Inventory (BFI) is done. To extract and select appropriate features for the classification, we employ an auto-encoder as a nonlinear feature learning technique. Since an auto-encoder does not extract proper classification lonely, a saddle point is found by a stop criterion based on maximum separate ability in binary classes. The results reveal that nonlinear features enhance the classification results in most personality traits. Furthermore, we use an adaptive neuro-fuzzy inference system classification to model the uncertainty rooted in mental states and affect the classification results through the extracted features. The classification outcomes on SSPNet Speaker Personality dataset demonstrate significant improvement in the results of four traits. These outgrowths verify the existence of uncertainty in the speech signal.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"35 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126302700","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420559
Abdolah Amirany, M. Meghdadi, M. H. Moaiyeri, Kian Jafari
The hardware implementation of neural network has always been of interest to the researchers as it can significantly increase the efficiency and application of neural networks due to the distributed nature of Artificial Neural Networks (ANNs) in both memory and computation. Direct implementation of ANNs also offer large gains when scaling the network sizes. Stochastic neurons are among the most significant aspects of machine learning algorithms and are very important in different neural networks. In this paper, a hardware model for the stochastic neuron based on the magnetic tunnel junction (MTJ) in subcritical current switching regime is proposed. Functional evaluation of the proposed model demonstrates that the behavior of the proposed model is comparable to the mathematical description of the stochastic neuron, and it has a negligible error in comparison with the theoretical model. The simulation results of image binarization over 10,000 images indicate that the proposed hardware model has only 0.25% pack signal to noise ratio (PSNR) and 0.02% structural similarity (SSIM) variation compared to its software-based counterpart.
{"title":"Stochastic Spintronic Neuron with Application to Image Binarization","authors":"Abdolah Amirany, M. Meghdadi, M. H. Moaiyeri, Kian Jafari","doi":"10.1109/CSICC52343.2021.9420559","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420559","url":null,"abstract":"The hardware implementation of neural network has always been of interest to the researchers as it can significantly increase the efficiency and application of neural networks due to the distributed nature of Artificial Neural Networks (ANNs) in both memory and computation. Direct implementation of ANNs also offer large gains when scaling the network sizes. Stochastic neurons are among the most significant aspects of machine learning algorithms and are very important in different neural networks. In this paper, a hardware model for the stochastic neuron based on the magnetic tunnel junction (MTJ) in subcritical current switching regime is proposed. Functional evaluation of the proposed model demonstrates that the behavior of the proposed model is comparable to the mathematical description of the stochastic neuron, and it has a negligible error in comparison with the theoretical model. The simulation results of image binarization over 10,000 images indicate that the proposed hardware model has only 0.25% pack signal to noise ratio (PSNR) and 0.02% structural similarity (SSIM) variation compared to its software-based counterpart.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"24 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123005440","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420621
Maryam Mahootiha, S. Golpayegani, B. Sadeghian
Money laundering nowadays occurs as one of the most severe and common crimes with great potential to harm the economy. Discovering money laundering by different computer methods has always been necessary due to criminals' high tendency to launder money. This study has focused on catching a type of money laundering, which leaves a trace in the datasets where the process of money laundering has been done collaboratively. This crime can be uncovered merely by discovering the pattern of group behavior of individuals. In this research, the social networks analysis method has been employed to detect group behavior in money laundering. The data were simulated based on the real environment and by considering different states because of proper data inaccessibility. The patterns of placement, layering, and integration of money are initially explained in money laundering in this study, followed by drawing a social network of individuals' transactions. In the end, the main culprits and their collaborators will be introduced based on a combination of criteria of centrality and detecting communities. Three different types of data have been used aimed at assessing the accuracy of the proposed solution. The proposed solution has also been compared with essential solutions such as the support vector machine, decision tree, and deep learning.
{"title":"Designing a New Method for Detecting Money Laundering based on Social Network Analysis","authors":"Maryam Mahootiha, S. Golpayegani, B. Sadeghian","doi":"10.1109/CSICC52343.2021.9420621","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420621","url":null,"abstract":"Money laundering nowadays occurs as one of the most severe and common crimes with great potential to harm the economy. Discovering money laundering by different computer methods has always been necessary due to criminals' high tendency to launder money. This study has focused on catching a type of money laundering, which leaves a trace in the datasets where the process of money laundering has been done collaboratively. This crime can be uncovered merely by discovering the pattern of group behavior of individuals. In this research, the social networks analysis method has been employed to detect group behavior in money laundering. The data were simulated based on the real environment and by considering different states because of proper data inaccessibility. The patterns of placement, layering, and integration of money are initially explained in money laundering in this study, followed by drawing a social network of individuals' transactions. In the end, the main culprits and their collaborators will be introduced based on a combination of criteria of centrality and detecting communities. Three different types of data have been used aimed at assessing the accuracy of the proposed solution. The proposed solution has also been compared with essential solutions such as the support vector machine, decision tree, and deep learning.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117096354","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420564
E. Tahanian, Mohsen Rezvani, Mansoor Fateh
Using Wireless Network-on-Chip (WiNOC) for designing the multicore System-on-Chips can significantly decrease the latency and power dissipation of the network. This improvement is achieved by replacing the multi-hop paths between far apart cores with a wireless single-hop link. Due to space, power and cost limitations, it is crucial to determine both the optimum number of equipped wireless hubs and their proper positions. In this paper, we propose a novel approach to obtain the optimum configuration of a WiNOC by leveraging the Simulating Annealing algorithm. Simultaneous multiple communications in such a network can be achieved by using multiple access techniques such as Orthogonal Frequency Division Multiple Access (OFDMA) to create dedicated channels between a source and destination pair. This technique is more bandwidth efficient compared to previously used FDMA. Also, it can distribute the available bandwidth between wireless nodes according to the traffic demands. So, we introduce an adequate channel reallocation algorithm regards to the broadcasting nature of the OFDMA scheme. The introduced architecture shows better performance in comparison with the conventional WiNOCs. This improvement is especially observed for latency characteristics where an improvement of about 15 is obtained.
{"title":"A Novel Wireless Network-on-Chip Architecture for Multicore Systems","authors":"E. Tahanian, Mohsen Rezvani, Mansoor Fateh","doi":"10.1109/CSICC52343.2021.9420564","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420564","url":null,"abstract":"Using Wireless Network-on-Chip (WiNOC) for designing the multicore System-on-Chips can significantly decrease the latency and power dissipation of the network. This improvement is achieved by replacing the multi-hop paths between far apart cores with a wireless single-hop link. Due to space, power and cost limitations, it is crucial to determine both the optimum number of equipped wireless hubs and their proper positions. In this paper, we propose a novel approach to obtain the optimum configuration of a WiNOC by leveraging the Simulating Annealing algorithm. Simultaneous multiple communications in such a network can be achieved by using multiple access techniques such as Orthogonal Frequency Division Multiple Access (OFDMA) to create dedicated channels between a source and destination pair. This technique is more bandwidth efficient compared to previously used FDMA. Also, it can distribute the available bandwidth between wireless nodes according to the traffic demands. So, we introduce an adequate channel reallocation algorithm regards to the broadcasting nature of the OFDMA scheme. The introduced architecture shows better performance in comparison with the conventional WiNOCs. This improvement is especially observed for latency characteristics where an improvement of about 15 is obtained.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115227167","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420540
Maryam Asadolahzade Kermanshahi, A. Akbari, B. Nasersharif
End-to-end models are state of the art for Automatic Speech Recognition (ASR) systems. Despite all their advantages, they suffer a significant problem: huge amounts of training data are required to achieve excellent performance. This problem is a serious challenge for low-resource languages such as Persian. Therefore, we need some methods and techniques to overcome this issue. One simple, yet effective method towards addressing this issue is transfer learning. We aim to explore the effect of transfer learning on a speech recognition system for the Persian language. To this end, we first train the network on 960 hours of English LibriSpeech corpus. Then, we transfer the trained network and fine-tune it on only about 3.5 hours of training data from the Persian FarsDat corpus. Transfer learning exhibits better performance while needing shorter training time than the model trained from scratch. Experimental results on FarsDat corpus indicate that transfer learning with a few hours of Persian training data can achieve 31.48% relative Phoneme Error Rate (PER) reduction compared to the model trained from scratch.
{"title":"Transfer Learning for End-to-End ASR to Deal with Low-Resource Problem in Persian Language","authors":"Maryam Asadolahzade Kermanshahi, A. Akbari, B. Nasersharif","doi":"10.1109/CSICC52343.2021.9420540","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420540","url":null,"abstract":"End-to-end models are state of the art for Automatic Speech Recognition (ASR) systems. Despite all their advantages, they suffer a significant problem: huge amounts of training data are required to achieve excellent performance. This problem is a serious challenge for low-resource languages such as Persian. Therefore, we need some methods and techniques to overcome this issue. One simple, yet effective method towards addressing this issue is transfer learning. We aim to explore the effect of transfer learning on a speech recognition system for the Persian language. To this end, we first train the network on 960 hours of English LibriSpeech corpus. Then, we transfer the trained network and fine-tune it on only about 3.5 hours of training data from the Persian FarsDat corpus. Transfer learning exhibits better performance while needing shorter training time than the model trained from scratch. Experimental results on FarsDat corpus indicate that transfer learning with a few hours of Persian training data can achieve 31.48% relative Phoneme Error Rate (PER) reduction compared to the model trained from scratch.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"29 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114584799","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 : 2021-03-03DOI: 10.1109/CSICC52343.2021.9420607
M. Attar, R. A. Moghadam, A. Rezaee
This paper presents a neuromechanical logic gate using Radio Frequency MEMS (RF MEMS) oscillators which are implemented as neurons of Hopfield network constituting an OR logic gate. Auto-correlative associative memory property being provided by phase-locked synchronized network of oscillators makes this logic operation possible. The proposed gate consists of 8 MEMS oscillators connected via electrical couplings and is capable of very high speed computation in case of utilizing high frequency MEMS resonators. This work can lay the groundwork for a new approach in analog computing systems based on mechanical oscillations.
{"title":"Design and Simulation of OR Logic Gate Using RF MEMS Resonators","authors":"M. Attar, R. A. Moghadam, A. Rezaee","doi":"10.1109/CSICC52343.2021.9420607","DOIUrl":"https://doi.org/10.1109/CSICC52343.2021.9420607","url":null,"abstract":"This paper presents a neuromechanical logic gate using Radio Frequency MEMS (RF MEMS) oscillators which are implemented as neurons of Hopfield network constituting an OR logic gate. Auto-correlative associative memory property being provided by phase-locked synchronized network of oscillators makes this logic operation possible. The proposed gate consists of 8 MEMS oscillators connected via electrical couplings and is capable of very high speed computation in case of utilizing high frequency MEMS resonators. This work can lay the groundwork for a new approach in analog computing systems based on mechanical oscillations.","PeriodicalId":374593,"journal":{"name":"2021 26th International Computer Conference, Computer Society of Iran (CSICC)","volume":"20 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115334048","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：