Pub Date : 2022-05-15DOI: 10.1109/SIU55565.2022.9864838
Abdulsamet Dagasan, Mustafa Akur, Mehmet Umut Demircin
Fiber Optic Distributed Acoustic Sensing (DAS) Systems use standard telecommunication fibers to detect acoustic vibrations up to 50 kms along the cable. In this paper we propose algorithms to detect fiber optic cable termination points and optical signal losses using DAS data. Proposed algorithms add traditional Optical Time-Domain Reflectometer (OTDR) measurement functionality to the DAS systems. Cable termination detection algorithm models the noise data in DAS signal that consists of electronic noise [e.g. Analog-to-Digital Converter (ADC) noises] and optical laser reflection noise. The cable termination detection algorithm analyzes noise statistics of the sensor data and finds the location where optic noise is no longer present. Signal loss detection algorithm first eliminates the environmental acoustic noise from the DAS signal; then, change point detection algorithm is applied to detect locations where significant signal loss occurs. Proposed algorithms are tested in various DAS installations in Turkey. Predicted cable termination and signal loss locations agree with OTDR measurements.
{"title":"Fiber Optic Cable Termination and Signal Loss Detection in DAS Systems","authors":"Abdulsamet Dagasan, Mustafa Akur, Mehmet Umut Demircin","doi":"10.1109/SIU55565.2022.9864838","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864838","url":null,"abstract":"Fiber Optic Distributed Acoustic Sensing (DAS) Systems use standard telecommunication fibers to detect acoustic vibrations up to 50 kms along the cable. In this paper we propose algorithms to detect fiber optic cable termination points and optical signal losses using DAS data. Proposed algorithms add traditional Optical Time-Domain Reflectometer (OTDR) measurement functionality to the DAS systems. Cable termination detection algorithm models the noise data in DAS signal that consists of electronic noise [e.g. Analog-to-Digital Converter (ADC) noises] and optical laser reflection noise. The cable termination detection algorithm analyzes noise statistics of the sensor data and finds the location where optic noise is no longer present. Signal loss detection algorithm first eliminates the environmental acoustic noise from the DAS signal; then, change point detection algorithm is applied to detect locations where significant signal loss occurs. Proposed algorithms are tested in various DAS installations in Turkey. Predicted cable termination and signal loss locations agree with OTDR measurements.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"74 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133388668","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864703
Erdem Bayhan, Cenk Berkan Deligoz, Feride Seymen, Mustafa Namdar, Arif Basgumus
In this study, the detection of the objects that they may encounter with deep learning models and the methods of the tactile paving surface tracking with Hough’s theorem are presented so that visually impaired individuals can easily walk outdoors. In the proposed approach, the training is primarily realized for machine learning of the deep learning models. The Faster R-CNN model and the SSD MobileNetV2 model are used in the training, and the accuracy performances of these two models are compared. During the training phase of the two models, a data set is generated using real-time and internet-based photographs. The training is completed by making use of 3653 photographs for 11 different objects that visually impaired individuals may encounter. In the detection of the objects, the accuracy rate of Faster R-CNN model is approximately 91%, and the SSD MobileNetV2 model achieved approximately 93% success. In addition, with the help of Hough’s theorem, it is observed that the edge surface lines are followed correctly in the tracking of the tactile paving surfaces.
{"title":"Unimpeded Walking with Deep Learning","authors":"Erdem Bayhan, Cenk Berkan Deligoz, Feride Seymen, Mustafa Namdar, Arif Basgumus","doi":"10.1109/SIU55565.2022.9864703","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864703","url":null,"abstract":"In this study, the detection of the objects that they may encounter with deep learning models and the methods of the tactile paving surface tracking with Hough’s theorem are presented so that visually impaired individuals can easily walk outdoors. In the proposed approach, the training is primarily realized for machine learning of the deep learning models. The Faster R-CNN model and the SSD MobileNetV2 model are used in the training, and the accuracy performances of these two models are compared. During the training phase of the two models, a data set is generated using real-time and internet-based photographs. The training is completed by making use of 3653 photographs for 11 different objects that visually impaired individuals may encounter. In the detection of the objects, the accuracy rate of Faster R-CNN model is approximately 91%, and the SSD MobileNetV2 model achieved approximately 93% success. In addition, with the help of Hough’s theorem, it is observed that the edge surface lines are followed correctly in the tracking of the tactile paving surfaces.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"153 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133795602","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864713
Dogukan Mesci, Anil Koluacik, B. Yılmaz, Melih Sen, E. Masazade, V. Beskardes
Bats are of great importance for the survival of all living beings and for biodiversity. This study aims to classify the collective calls of the Egyptian fruit bat, whose northernmost distribution is in Turkey, using deep learning methods CNN and LSTM and utilizing MFCC (Mel Frequency Cepstral Coefficients) features. Thanks to the classification of species-specific calls, it is possible to observe the habitat preference, social relations, foraging, reproduction, mobility and migration of the species. The classification results obtained in this study provide significant increases compared to the previous study, especially in distinguishing certain calls.
蝙蝠对所有生物的生存和生物多样性至关重要。本研究旨在利用深度学习方法CNN和LSTM，利用Mel Frequency Cepstral Coefficients特征，对分布在土耳其最北的埃及果蝠的集体鸣叫进行分类。由于对物种特有的叫声进行了分类，因此有可能观察到物种的栖息地偏好、社会关系、觅食、繁殖、迁移和迁徙。与以往的研究相比，本研究获得的分类结果有了显著的提高，特别是在区分某些叫声方面。
{"title":"Classification of Egyptian Fruit Bat Calls with Deep Learning Methods","authors":"Dogukan Mesci, Anil Koluacik, B. Yılmaz, Melih Sen, E. Masazade, V. Beskardes","doi":"10.1109/SIU55565.2022.9864713","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864713","url":null,"abstract":"Bats are of great importance for the survival of all living beings and for biodiversity. This study aims to classify the collective calls of the Egyptian fruit bat, whose northernmost distribution is in Turkey, using deep learning methods CNN and LSTM and utilizing MFCC (Mel Frequency Cepstral Coefficients) features. Thanks to the classification of species-specific calls, it is possible to observe the habitat preference, social relations, foraging, reproduction, mobility and migration of the species. The classification results obtained in this study provide significant increases compared to the previous study, especially in distinguishing certain calls.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"9 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115364526","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864689
Tuğba Erkoç, M. T. Eskil
Effective training of filters in Convolutional Neural Networks (CNN) ensures their success. In order to achieve good classification results in CNNs, filters must be carefully initialized, trained and fine-tuned. We propose an unsupervised method that allows the discovery of filters from the given dataset in a single epoch without specifying the number of filters hyper-parameter in convolutional layers. Our proposed method gradually builds the convolutional layers by a discovery routine that extracts a number of features that adequately represent the complexity of the input domain. The discovered filters represent the patterns in the domain, so they do not require any initialization method or backpropagation training for fine tuning purposes. Our method achieves 99.03% accuracy on MNIST dataset without applying any data augmentation techniques.
{"title":"Unsupervised Similarity Based Convolutions for Handwritten Digit Classification","authors":"Tuğba Erkoç, M. T. Eskil","doi":"10.1109/SIU55565.2022.9864689","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864689","url":null,"abstract":"Effective training of filters in Convolutional Neural Networks (CNN) ensures their success. In order to achieve good classification results in CNNs, filters must be carefully initialized, trained and fine-tuned. We propose an unsupervised method that allows the discovery of filters from the given dataset in a single epoch without specifying the number of filters hyper-parameter in convolutional layers. Our proposed method gradually builds the convolutional layers by a discovery routine that extracts a number of features that adequately represent the complexity of the input domain. The discovered filters represent the patterns in the domain, so they do not require any initialization method or backpropagation training for fine tuning purposes. Our method achieves 99.03% accuracy on MNIST dataset without applying any data augmentation techniques.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"26 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114694672","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864956
Ugur Erbas, M. Tabakcioglu
With the developing communication technology in recent years, the importance of placing the base stations in the right location has increased in order to ensure a healthy communication. It is thought that this situation will become even more important with 5G technology. In this study, 2D maps with earth maps and transformation windows were created in MATLAB using 3D digital data. The diffracted, direct and reflected rays were determined, and the ray tracing algorithm was run for the superconducting surface. A 3D coverage area is mapped for a possible transmitter position. Electric field graphs are drawn for different heights. It has been observed that the electric field graph changes depending on the landforms, distance, diffraction and interference of the rays.
{"title":"Generation of 3D Coverage Map","authors":"Ugur Erbas, M. Tabakcioglu","doi":"10.1109/SIU55565.2022.9864956","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864956","url":null,"abstract":"With the developing communication technology in recent years, the importance of placing the base stations in the right location has increased in order to ensure a healthy communication. It is thought that this situation will become even more important with 5G technology. In this study, 2D maps with earth maps and transformation windows were created in MATLAB using 3D digital data. The diffracted, direct and reflected rays were determined, and the ray tracing algorithm was run for the superconducting surface. A 3D coverage area is mapped for a possible transmitter position. Electric field graphs are drawn for different heights. It has been observed that the electric field graph changes depending on the landforms, distance, diffraction and interference of the rays.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114250365","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864987
Ahmet Faruk Celimli, Ogulcan Özdemir, L. Akarun
Sign languages are visual languages expressed with multiple cues including facial expressions, upper-body and hand gestures. These different visual cues can be used together or at different instants to convey the message. In order to recognize sign languages, it is crucial to model what, where and when to attend. In this study, we developed a model to use different visual cues at the same time by using Temporal Shift Modules (TSMs) and attention modeling. Our experiments are conducted with BospohorusSign22k dataset. Our system has achieved 92.46% recognition accuracy and improved the performance approximately 14% compared to the baseline study with 78.85% accuracy.
{"title":"Attention Modeling with Temporal Shift in Sign Language Recognition","authors":"Ahmet Faruk Celimli, Ogulcan Özdemir, L. Akarun","doi":"10.1109/SIU55565.2022.9864987","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864987","url":null,"abstract":"Sign languages are visual languages expressed with multiple cues including facial expressions, upper-body and hand gestures. These different visual cues can be used together or at different instants to convey the message. In order to recognize sign languages, it is crucial to model what, where and when to attend. In this study, we developed a model to use different visual cues at the same time by using Temporal Shift Modules (TSMs) and attention modeling. Our experiments are conducted with BospohorusSign22k dataset. Our system has achieved 92.46% recognition accuracy and improved the performance approximately 14% compared to the baseline study with 78.85% accuracy.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"284 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116101742","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864728
Ö. Mercan, Umut Özdil, Sükrü Ozan
This study was carried out with the aim of automatically translating phone calls between customers and customer representatives of a company. The dataset used in the study was created with audio files that were taken from open source platforms and reading of short texts in various contents by the company personnel. In addition to the labbeled data, approximately 28 thousand unlabeled data were labelled, and a total of 37534 audio data were prepared to be used in the training of the model that will translate from speech to text. The Wav2Vec2-XLSR-53 model which is a pre-trained model trained in 53 languages was fine-tuned with the our Turkish dataset. It has been obtained that it gives successful results in the speech to text performed on the data that is not used in model training and validation. The model was shared as open source on HugginFace to be used and tested for similar speech to text translation problems.
{"title":"Çok Dilli Sesten Metne Çeviri Modelinin İnce Ayar Yapılarak Türkçe Dilindeki Başarısının Arttırılması Increasing Performance in Turkish by Finetuning of Multilingual Speech-to-Text Model","authors":"Ö. Mercan, Umut Özdil, Sükrü Ozan","doi":"10.1109/SIU55565.2022.9864728","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864728","url":null,"abstract":"This study was carried out with the aim of automatically translating phone calls between customers and customer representatives of a company. The dataset used in the study was created with audio files that were taken from open source platforms and reading of short texts in various contents by the company personnel. In addition to the labbeled data, approximately 28 thousand unlabeled data were labelled, and a total of 37534 audio data were prepared to be used in the training of the model that will translate from speech to text. The Wav2Vec2-XLSR-53 model which is a pre-trained model trained in 53 languages was fine-tuned with the our Turkish dataset. It has been obtained that it gives successful results in the speech to text performed on the data that is not used in model training and validation. The model was shared as open source on HugginFace to be used and tested for similar speech to text translation problems.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"32 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121815103","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864775
Ismail Can Büyüktepe, A. K. Hocaoglu
In this study, an algorithm that can classify human and car has been developed by using vibration signals obtained from a three-axis accelerometer sensor station placed on three different floors. Data were collected over soil, asphalt and concrete ground. As classifiers, k-Nearest Neighbor classifier (k-NN) and Support Vector Machine (SVM) classifiers are used. Using classifiers alone limits classification performance. A two-stage classifier model has been proposed to improve the classification performance. The classifier model, which is proposed in two stages, detects the presence of motion in the first stage. In the second stage, it performs the classification of moving targets. As a result of the experimental studies, it has been shown that the proposed two-stage classifier model improves the performance in solving the problem.
{"title":"Classification of Moving Ground Targets Using Measurement from Accelerometer on Road Surface","authors":"Ismail Can Büyüktepe, A. K. Hocaoglu","doi":"10.1109/SIU55565.2022.9864775","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864775","url":null,"abstract":"In this study, an algorithm that can classify human and car has been developed by using vibration signals obtained from a three-axis accelerometer sensor station placed on three different floors. Data were collected over soil, asphalt and concrete ground. As classifiers, k-Nearest Neighbor classifier (k-NN) and Support Vector Machine (SVM) classifiers are used. Using classifiers alone limits classification performance. A two-stage classifier model has been proposed to improve the classification performance. The classifier model, which is proposed in two stages, detects the presence of motion in the first stage. In the second stage, it performs the classification of moving targets. As a result of the experimental studies, it has been shown that the proposed two-stage classifier model improves the performance in solving the problem.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"166 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121480785","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864758
Tolga Tüfekçi, Oguz Ülgen, Serhat Erküçük, T. Baykaş
In order to satisfy the need for high data rate and high number of users, new generation communication techniques are developed. One of the techniques that may be used in future generation communication networks is Sparse Code Multiple Access (SCMA). With this new technique, the aim is to allocate users frequency resources in a non-orthogonal way by using code books. For this new technique, which is has a potential to be used in 5G and beyond communication networks, most researches have focused on flat fading channels and related results have been provided. In this work, different from earlier studies, fast fading channels have been considered for channels varying at different rates, and bit-error performance results have been provided with computer simulations.
{"title":"Performance of SCMA Systems in Fast-Fading Channels","authors":"Tolga Tüfekçi, Oguz Ülgen, Serhat Erküçük, T. Baykaş","doi":"10.1109/SIU55565.2022.9864758","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864758","url":null,"abstract":"In order to satisfy the need for high data rate and high number of users, new generation communication techniques are developed. One of the techniques that may be used in future generation communication networks is Sparse Code Multiple Access (SCMA). With this new technique, the aim is to allocate users frequency resources in a non-orthogonal way by using code books. For this new technique, which is has a potential to be used in 5G and beyond communication networks, most researches have focused on flat fading channels and related results have been provided. In this work, different from earlier studies, fast fading channels have been considered for channels varying at different rates, and bit-error performance results have been provided with computer simulations.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121072790","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 : 2022-05-15DOI: 10.1109/SIU55565.2022.9864848
E. Dogan, H. F. Ugurdag, Hasan Unlu
Applications of artificial neural networks on low-cost embedded systems and microcontrollers (MCUs), has recently been attracting more attention than ever. Since MCUs have limited memory capacity as well as limited compute-speed compared to workstations, employment of current deep learning algorithms on MCUs becomes more practical with the help of model compression. This makes MCUs common and practical alternative solution for autonomous systems. In this paper, we add model compression, specifically Deep Compression, to an existing work, which efficiently deploys PyTorch models on MCUs, in order to increase neural network speed and save electrical power. First, we prune the weight values close to zero in convolutional and fully connected layers. Secondly, the remaining weights and activations are quantized to 8-bit integers from 32-bit floating-point. Finally, forward pass functions are compressed using special data structures for sparse matrices, which store only nonzero weights. In the case of the LeNet-5 model, the memory footprint was reduced by 12.5x, and the inference speed was boosted by 2.6x.
{"title":"Using Deep Compression on PyTorch Models for Autonomous Systems","authors":"E. Dogan, H. F. Ugurdag, Hasan Unlu","doi":"10.1109/SIU55565.2022.9864848","DOIUrl":"https://doi.org/10.1109/SIU55565.2022.9864848","url":null,"abstract":"Applications of artificial neural networks on low-cost embedded systems and microcontrollers (MCUs), has recently been attracting more attention than ever. Since MCUs have limited memory capacity as well as limited compute-speed compared to workstations, employment of current deep learning algorithms on MCUs becomes more practical with the help of model compression. This makes MCUs common and practical alternative solution for autonomous systems. In this paper, we add model compression, specifically Deep Compression, to an existing work, which efficiently deploys PyTorch models on MCUs, in order to increase neural network speed and save electrical power. First, we prune the weight values close to zero in convolutional and fully connected layers. Secondly, the remaining weights and activations are quantized to 8-bit integers from 32-bit floating-point. Finally, forward pass functions are compressed using special data structures for sparse matrices, which store only nonzero weights. In the case of the LeNet-5 model, the memory footprint was reduced by 12.5x, and the inference speed was boosted by 2.6x.","PeriodicalId":115446,"journal":{"name":"2022 30th Signal Processing and Communications Applications Conference (SIU)","volume":"16 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122852364","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：