Pub Date : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00207
Louis Jensen, Jayme Fosa, Ben Teitelbaum, Peter Chin
Time series data has become ubiquitous in the modern era of data collection. With the increase of these time series data streams, the demand for automatic time series anomaly detection has also increased. Automatic monitoring of data allows engineers to investigate only unusual behavior in their data streams. Despite this increase in demand for automatic time series anomaly detection, many popular methods fail to offer a general purpose solution. Some demand expensive labelling of anomalies, others require the data to follow certain assumed patterns, some have long and unstable training, and many suffer from high rates of false alarms. In this paper we demonstrate that simpler is often better, showing that a fully unsupervised multilayer perceptron autoencoder is able to outperform much more complicated models with only a few critical improvements. We offer improvements to help distinguish anomalous subsequences near to each other, and to distinguish anomalies even in the midst of changing distributions of data. We compare our model with state-of-the-art competitors on benchmark datasets sourced from NASA, Yahoo, and Numenta, achieving improvements beyond competitive models in all three datasets.
{"title":"How Dense Autoencoders can still Achieve the State-of-the-art in Time-Series Anomaly Detection","authors":"Louis Jensen, Jayme Fosa, Ben Teitelbaum, Peter Chin","doi":"10.1109/ICMLA52953.2021.00207","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00207","url":null,"abstract":"Time series data has become ubiquitous in the modern era of data collection. With the increase of these time series data streams, the demand for automatic time series anomaly detection has also increased. Automatic monitoring of data allows engineers to investigate only unusual behavior in their data streams. Despite this increase in demand for automatic time series anomaly detection, many popular methods fail to offer a general purpose solution. Some demand expensive labelling of anomalies, others require the data to follow certain assumed patterns, some have long and unstable training, and many suffer from high rates of false alarms. In this paper we demonstrate that simpler is often better, showing that a fully unsupervised multilayer perceptron autoencoder is able to outperform much more complicated models with only a few critical improvements. We offer improvements to help distinguish anomalous subsequences near to each other, and to distinguish anomalies even in the midst of changing distributions of data. We compare our model with state-of-the-art competitors on benchmark datasets sourced from NASA, Yahoo, and Numenta, achieving improvements beyond competitive models in all three datasets.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"1 1","pages":"1272-1277"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"82949939","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}
In club football, scouting is a crucial aspect of player recruitment, with elite football clubs investing millions of dollars in scouting and signing the best player for their team every year. Scouting requires great analytical and observational skills from the scout, to find the best player for any position in the team. A scout needs to analyze the player by watching his in-game actions, physical attributes and make a judgement on how the player might fit into the team. Every team has a formation, a style of play and a specific profile of player is required for a given position depending on the aforementioned factors. But scouts only watch a player play a few matches in person, and prepare their scouting report based on a player’s performance in those matches. This process is flawed as the scout is expected to watch a few games and make estimates of the player’s performance in a new team. The player statistics can help the scout in making better data-driven decisions. A player’s career statistics can provide a picture of how the player performs individually, but they fail to predict player chemistry alongside a team. Misjudgement in scouting can lead to losses of millions of dollars to a club. We propose to solve this problem by utilising vast amounts of quantitative and qualitative player statistics (from 3+ sources), and by incorporating data science and machine learning algorithms to simulate real world performances of the team after the addition of the newly scouted player. We take into account specific player requirements and classify a player into one of our specific 15 player types, and use the team’s formation and style of play to predict the players that will have the best chemistry with any given lineup, thereby facilitating scouts in making better decisions.
{"title":"Data Driven football scouting assistance with simulated player performance extrapolation","authors":"Shantanu Ghar, Sayali Patil, Venkhatesh Arunachalam","doi":"10.1109/ICMLA52953.2021.00189","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00189","url":null,"abstract":"In club football, scouting is a crucial aspect of player recruitment, with elite football clubs investing millions of dollars in scouting and signing the best player for their team every year. Scouting requires great analytical and observational skills from the scout, to find the best player for any position in the team. A scout needs to analyze the player by watching his in-game actions, physical attributes and make a judgement on how the player might fit into the team. Every team has a formation, a style of play and a specific profile of player is required for a given position depending on the aforementioned factors. But scouts only watch a player play a few matches in person, and prepare their scouting report based on a player’s performance in those matches. This process is flawed as the scout is expected to watch a few games and make estimates of the player’s performance in a new team. The player statistics can help the scout in making better data-driven decisions. A player’s career statistics can provide a picture of how the player performs individually, but they fail to predict player chemistry alongside a team. Misjudgement in scouting can lead to losses of millions of dollars to a club. We propose to solve this problem by utilising vast amounts of quantitative and qualitative player statistics (from 3+ sources), and by incorporating data science and machine learning algorithms to simulate real world performances of the team after the addition of the newly scouted player. We take into account specific player requirements and classify a player into one of our specific 15 player types, and use the team’s formation and style of play to predict the players that will have the best chemistry with any given lineup, thereby facilitating scouts in making better decisions.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"19 1","pages":"1160-1167"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"84331322","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00170
D. Schwartz, Jonathan Michael Gomes Selman, P. Wrege, A. Paepcke
Artificial intelligence is increasingly used in ecological contexts to monitor animal and insect populations. Species of interest are those in danger of extinction, and those that play pivotal roles in agriculture. Noticing population declines or geographical shifts early enough for intervention can prevent local famine and disruption to the global food chain. Traditionally, data are collected in the field using human labor or sensors. Applicable classification models then analyze the data on central servers. The most expensive, and sometimes dangerous part of the remote sensing solution is the human labor of visiting the sensors, retrieving data, and changing batteries. Constantly sending all readings by radio is expensive in power. Instead, having AI in the sensors process readings, and only transmitting results could lead to an indefinitely autonomous, renewably powered solution. We implemented an elephant vocalization detector on a small processor board, and demonstrate that such a device can be operated at low enough power levels with considerable freedom of choice among AI technologies. We achieved a mean of 1.6W, in the best case staying within 75% of memory limits. Measurements covered three inference models, two batch sizes, and two floating point word width settings.
{"title":"Deployment of Embedded Edge-AI for Wildlife Monitoring in Remote Regions","authors":"D. Schwartz, Jonathan Michael Gomes Selman, P. Wrege, A. Paepcke","doi":"10.1109/ICMLA52953.2021.00170","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00170","url":null,"abstract":"Artificial intelligence is increasingly used in ecological contexts to monitor animal and insect populations. Species of interest are those in danger of extinction, and those that play pivotal roles in agriculture. Noticing population declines or geographical shifts early enough for intervention can prevent local famine and disruption to the global food chain. Traditionally, data are collected in the field using human labor or sensors. Applicable classification models then analyze the data on central servers. The most expensive, and sometimes dangerous part of the remote sensing solution is the human labor of visiting the sensors, retrieving data, and changing batteries. Constantly sending all readings by radio is expensive in power. Instead, having AI in the sensors process readings, and only transmitting results could lead to an indefinitely autonomous, renewably powered solution. We implemented an elephant vocalization detector on a small processor board, and demonstrate that such a device can be operated at low enough power levels with considerable freedom of choice among AI technologies. We achieved a mean of 1.6W, in the best case staying within 75% of memory limits. Measurements covered three inference models, two batch sizes, and two floating point word width settings.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"1 1","pages":"1035-1042"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"89243102","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00098
Patrick Bujok, Maria Jensen, Steffen M. Larsen, R. A. Alphinas
Machine Learning applications provide a promising method to support clinical practitioners in Breast Cancer (BC) detection. Currently, Fine Needle Aspiration (FNA) is a commonly applied diagnostic method for BC tumors, which, however, is associated with ominous false negative misclassifications. For this purpose, the present study explores Artificial Neural Networks (ANNs) with the aim of outperforming clinical practices via FNA in classifying benign or malignant BC cases with regard to an improved accuracy and reduced False Negative Rate (FNR) using the Breast Cancer Wisconsin (Diagnostic) Dataset (WDBC). The findings reveal that a dense ANN with a single hidden layer including 15 neurons can reach a testing accuracy of 98.60% and a FNR of 0% on a scaled dataset. In combination with several introduced improvement measures, a high degree of generalizability is associated with the model under the consideration of the relatively small dataset. As a result, this model outperforms not only clinical practitioners but also 72 classifiers from the recent literature.
{"title":"Outperforming Clinical Practices in Breast Cancer Detection: A Superior Dense Neural Network in Classification and False Negative Reduction","authors":"Patrick Bujok, Maria Jensen, Steffen M. Larsen, R. A. Alphinas","doi":"10.1109/ICMLA52953.2021.00098","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00098","url":null,"abstract":"Machine Learning applications provide a promising method to support clinical practitioners in Breast Cancer (BC) detection. Currently, Fine Needle Aspiration (FNA) is a commonly applied diagnostic method for BC tumors, which, however, is associated with ominous false negative misclassifications. For this purpose, the present study explores Artificial Neural Networks (ANNs) with the aim of outperforming clinical practices via FNA in classifying benign or malignant BC cases with regard to an improved accuracy and reduced False Negative Rate (FNR) using the Breast Cancer Wisconsin (Diagnostic) Dataset (WDBC). The findings reveal that a dense ANN with a single hidden layer including 15 neurons can reach a testing accuracy of 98.60% and a FNR of 0% on a scaled dataset. In combination with several introduced improvement measures, a high degree of generalizability is associated with the model under the consideration of the relatively small dataset. As a result, this model outperforms not only clinical practitioners but also 72 classifiers from the recent literature.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"25 1","pages":"589-594"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"87677746","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00167
A. Hafeez, Eduardo Alonso, Aram Ter-Sarkisov
Predictive maintenance, which has traditionally used anomaly detection methods on sensory data, is now being replaced by event-based techniques. These methods utilise events with multiple temporal (and often non-numeric) features, produced by diagnostic modules. This raises the need of learning numerical event representations to predict the next fault event in industrial machines, specially vehicles, that use Diagnostic Trouble Codes (DTCs). We propose a predictive maintenance approach, named Sequential Multivariate Fault Prediction (SMFP), for predicting the next multivariate DTC fault in an event sequence, using Long Short-Term Memory Networks (LSTMs) and jointly learned event embeddings. By performing an in-depth comparison of different architectural choices and contextual preprocessing techniques, we provide an initial baseline for SMFP that achieves top-3 accuracy of 63% on predicting multivariate fault with 3 collective output layers, using vehicle maintenance data as a case study.
{"title":"Towards Sequential Multivariate Fault Prediction for Vehicular Predictive Maintenance","authors":"A. Hafeez, Eduardo Alonso, Aram Ter-Sarkisov","doi":"10.1109/ICMLA52953.2021.00167","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00167","url":null,"abstract":"Predictive maintenance, which has traditionally used anomaly detection methods on sensory data, is now being replaced by event-based techniques. These methods utilise events with multiple temporal (and often non-numeric) features, produced by diagnostic modules. This raises the need of learning numerical event representations to predict the next fault event in industrial machines, specially vehicles, that use Diagnostic Trouble Codes (DTCs). We propose a predictive maintenance approach, named Sequential Multivariate Fault Prediction (SMFP), for predicting the next multivariate DTC fault in an event sequence, using Long Short-Term Memory Networks (LSTMs) and jointly learned event embeddings. By performing an in-depth comparison of different architectural choices and contextual preprocessing techniques, we provide an initial baseline for SMFP that achieves top-3 accuracy of 63% on predicting multivariate fault with 3 collective output layers, using vehicle maintenance data as a case study.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"309 1","pages":"1016-1021"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"75682546","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00014
Aashish Rai, Clara Ducher, J. Cooperstock
With the recent popularization of generative frameworks for producing photorealistic face images, we now have the ability to create a convincing graphical match for any particular individual. It is unrealistic, however, to rely solely on such generative methods to randomly produce the facial characteristics we are seeking. Instead, manipulation of facial attributes in the latent space, enabled by the InterFaceGAN framework, allows us to “tweak” these characteristics in the desired direction to improve the quality of the match. The challenge in this process is that attribute entanglement leads to a change of one feature having an undesirable impact on others. We explore several strategies to improve the results of these manipulations, and demonstrate how the automatic conditioning of attributes can be used to minimize the impact of such entanglement, and further, allow for improved control over complex (non-binary) attributes such as race or face shape.
{"title":"Improved Attribute Manipulation in the Latent Space of StyleGAN for Semantic Face Editing","authors":"Aashish Rai, Clara Ducher, J. Cooperstock","doi":"10.1109/ICMLA52953.2021.00014","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00014","url":null,"abstract":"With the recent popularization of generative frameworks for producing photorealistic face images, we now have the ability to create a convincing graphical match for any particular individual. It is unrealistic, however, to rely solely on such generative methods to randomly produce the facial characteristics we are seeking. Instead, manipulation of facial attributes in the latent space, enabled by the InterFaceGAN framework, allows us to “tweak” these characteristics in the desired direction to improve the quality of the match. The challenge in this process is that attribute entanglement leads to a change of one feature having an undesirable impact on others. We explore several strategies to improve the results of these manipulations, and demonstrate how the automatic conditioning of attributes can be used to minimize the impact of such entanglement, and further, allow for improved control over complex (non-binary) attributes such as race or face shape.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"21 39","pages":"38-43"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"91505996","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00070
L. Dheekollu, H. Wadhwa, Siddharth Vimal, Anubhav Gupta, Siddhartha Asthana, Ankur Arora, Smriti Gupta
Predicting customer attrition (churn) is a well known problem in industries that provide services, like financial institutions, telecommunications, e-commerce, and retail. There are two kinds of attrition - active and passive (silent). Active attrition is usually associated with subscription-based business models, commonly seen in telecommunications and internet industries like Netflix. In industries like finance, retail, and ecommerce, we see the other kind of attrition - silent attrition where customers stop doing business without formal notice. This makes the silent attrition prediction problem even more challenging because it is difficult to differentiate between attrited and inactive customers. We focus our work on predicting silent attrition which is still under-explored in the payment card industry (i.e. Mastercard, Visa). The contribution of our work is threefold. First, we present a data-driven approach to define silent attrition as customer inactivity. Second, we discussed multiple procedures to generate synthetic data thereby preserving customers’ privacy. At last, we presented a comprehensive view of various machine learning (ML) pathways in which this churn prediction problem can be framed and solved; each requiring a specific feature engineering. We presented experimental results corresponding to each pathway to comparative analysis. We believe that this work to be beneficial to the researchers and ML practitioners who often have to deal with sensitive financial data but have limited permission to use it. In this direction, we demonstrated the use of synthetic data generation to reduce the risk of data leakage and other privacy concerns relating to ML models development.
{"title":"Modeling approaches for Silent Attrition prediction in Payment networks","authors":"L. Dheekollu, H. Wadhwa, Siddharth Vimal, Anubhav Gupta, Siddhartha Asthana, Ankur Arora, Smriti Gupta","doi":"10.1109/ICMLA52953.2021.00070","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00070","url":null,"abstract":"Predicting customer attrition (churn) is a well known problem in industries that provide services, like financial institutions, telecommunications, e-commerce, and retail. There are two kinds of attrition - active and passive (silent). Active attrition is usually associated with subscription-based business models, commonly seen in telecommunications and internet industries like Netflix. In industries like finance, retail, and ecommerce, we see the other kind of attrition - silent attrition where customers stop doing business without formal notice. This makes the silent attrition prediction problem even more challenging because it is difficult to differentiate between attrited and inactive customers. We focus our work on predicting silent attrition which is still under-explored in the payment card industry (i.e. Mastercard, Visa). The contribution of our work is threefold. First, we present a data-driven approach to define silent attrition as customer inactivity. Second, we discussed multiple procedures to generate synthetic data thereby preserving customers’ privacy. At last, we presented a comprehensive view of various machine learning (ML) pathways in which this churn prediction problem can be framed and solved; each requiring a specific feature engineering. We presented experimental results corresponding to each pathway to comparative analysis. We believe that this work to be beneficial to the researchers and ML practitioners who often have to deal with sensitive financial data but have limited permission to use it. In this direction, we demonstrated the use of synthetic data generation to reduce the risk of data leakage and other privacy concerns relating to ML models development.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"64 1","pages":"409-414"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"91086362","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00121
Wenjun Zeng, Bakhtiar Amen
Automated flower detection and control is important to crop production and precision agriculture. Some computer vision methods have been proposed for flower detection, but their performances are not satisfactory on platforms with limited computing ability such as mobile and embedded devices, and thus not suitable for field applications. Herein we demonstrate two de novo approaches that can precisely detect the flowers of two bioenergy crops (potatoes and sweet potatoes) and can distinguish them from similar flowers of relative species (eggplants and Ipomoea triloba) on mobile devices. In this work, a custom dataset containing 495 manually labelled images is constructed for training and testing, and the latest state-of-the-art object detection model, YOLOv4, as well as its lightweight version, YOLOv4-tiny, are selected as the flower detection models. Some other milestone object detection models including YOLOv3, YOLOv3-tiny, SSD and Faster-RCNN are chosen as benchmarks for performance comparison. The comparative experiment results indicate that the retrained YOLOv4 model achieves a considerable high mean average precision (mAP= 91%;) but a slower inference speed (FPS) on a mobile device, while the retrained YOLOv4-tiny has a lower mAP of 87%; but reach a higher FPS of 9 on a mobile device. Two mobile applications are then developed by directly deploying YOLOv4-tiny model on a mobile app and by deploying YOLOv4 on a web API, respectively. The testing experiments indicate that both applications can not only achieve real-time and accurate detection, but also reduce computation burdens on mobile devices.
{"title":"Applications of Mobile Machine Learning for Detecting Bio-energy Crops Flowers","authors":"Wenjun Zeng, Bakhtiar Amen","doi":"10.1109/ICMLA52953.2021.00121","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00121","url":null,"abstract":"Automated flower detection and control is important to crop production and precision agriculture. Some computer vision methods have been proposed for flower detection, but their performances are not satisfactory on platforms with limited computing ability such as mobile and embedded devices, and thus not suitable for field applications. Herein we demonstrate two de novo approaches that can precisely detect the flowers of two bioenergy crops (potatoes and sweet potatoes) and can distinguish them from similar flowers of relative species (eggplants and Ipomoea triloba) on mobile devices. In this work, a custom dataset containing 495 manually labelled images is constructed for training and testing, and the latest state-of-the-art object detection model, YOLOv4, as well as its lightweight version, YOLOv4-tiny, are selected as the flower detection models. Some other milestone object detection models including YOLOv3, YOLOv3-tiny, SSD and Faster-RCNN are chosen as benchmarks for performance comparison. The comparative experiment results indicate that the retrained YOLOv4 model achieves a considerable high mean average precision (mAP= 91%;) but a slower inference speed (FPS) on a mobile device, while the retrained YOLOv4-tiny has a lower mAP of 87%; but reach a higher FPS of 9 on a mobile device. Two mobile applications are then developed by directly deploying YOLOv4-tiny model on a mobile app and by deploying YOLOv4 on a web API, respectively. The testing experiments indicate that both applications can not only achieve real-time and accurate detection, but also reduce computation burdens on mobile devices.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"41 1","pages":"724-729"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"91011208","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00180
Tyler Cody, Stephen C. Adams, P. Beling
Metric learning is a well-developed field in machine learning and has seen recent application in the area of prognostics and health management (PHM). Metric learning allows for fault diagnosis or condition monitoring models to be developed with the assumption that a machine- or load-specific similarity metric can be learned after model deployment. Existing literature has used metric learning to fine-tune deep learning models to address machine-to-machine differences and differences in working conditions. Here, we study metric learning in isolation, not as an intermediate step in deep learning, by conducting a comparative study of Principal Component Analysis (PCA), Neighborhood Component Analysis (NCA), Local Fisher Discriminant Analysis (LFDA), and Large Margin Nearest Neighbor (LMNN). We consider performance metrics for prediction performance, cluster performance, feature sensitivity, sample efficiency, and latent space efficiency. We find that linear partitions on the latent spaces learned via metric learning are able to achieve accuracies greater than 90% on Case Western Reserve University’s bearing fault data set using only the drive-end vibration signal. We find PCA to be dominated by metric learning algorithms for all working loads considered. And, in sum, we demonstrate classical metric learning algorithms to be a promising approach for learning machine-and load-specific similarity metrics for PHM with minor data processing and small samples.
{"title":"Trade-offs in Metric Learning for Bearing Fault Diagnosis","authors":"Tyler Cody, Stephen C. Adams, P. Beling","doi":"10.1109/ICMLA52953.2021.00180","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00180","url":null,"abstract":"Metric learning is a well-developed field in machine learning and has seen recent application in the area of prognostics and health management (PHM). Metric learning allows for fault diagnosis or condition monitoring models to be developed with the assumption that a machine- or load-specific similarity metric can be learned after model deployment. Existing literature has used metric learning to fine-tune deep learning models to address machine-to-machine differences and differences in working conditions. Here, we study metric learning in isolation, not as an intermediate step in deep learning, by conducting a comparative study of Principal Component Analysis (PCA), Neighborhood Component Analysis (NCA), Local Fisher Discriminant Analysis (LFDA), and Large Margin Nearest Neighbor (LMNN). We consider performance metrics for prediction performance, cluster performance, feature sensitivity, sample efficiency, and latent space efficiency. We find that linear partitions on the latent spaces learned via metric learning are able to achieve accuracies greater than 90% on Case Western Reserve University’s bearing fault data set using only the drive-end vibration signal. We find PCA to be dominated by metric learning algorithms for all working loads considered. And, in sum, we demonstrate classical metric learning algorithms to be a promising approach for learning machine-and load-specific similarity metrics for PHM with minor data processing and small samples.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"23 1","pages":"1100-1105"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"82140624","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 : 2021-12-01DOI: 10.1109/ICMLA52953.2021.00088
Cagri Ozdemir, R. Hoover, Kyle A. Caudle
The tensor singular value decomposition (t-SVD) based on an algebra of circulants is an effective multilinear sub- space learning technique for dimensionality reduction and data classification. Unfortunately, the computational cost associated with computing the t-SVD can become prohibitively expensive, particularly when dealing with very large data sets. In this paper, we present a computationally efficient approach for estimating the t-SVD by capitalizing on the correlations of the data in the temporal dimension. The approach proceeds by extending our prior work on fast eigenspace decompositions by transforming the tensor data from the spatial domain to the spectral domain in order to obtain reduced order harmonic tensor. The t-SVD can then be applied in the transform domain thereby significantly reducing the computational burden. Experimental results which are presented on the extended Yale-B, COIL-100, and MNIST data sets show the proposed method provides considerable computational savings with the approximated subspaces that are nearly the same as the true subspaces as computed via the t-SVD.
{"title":"Fast Tensor Singular Value Decomposition Using the Low-Resolution Features of Tensors","authors":"Cagri Ozdemir, R. Hoover, Kyle A. Caudle","doi":"10.1109/ICMLA52953.2021.00088","DOIUrl":"https://doi.org/10.1109/ICMLA52953.2021.00088","url":null,"abstract":"The tensor singular value decomposition (t-SVD) based on an algebra of circulants is an effective multilinear sub- space learning technique for dimensionality reduction and data classification. Unfortunately, the computational cost associated with computing the t-SVD can become prohibitively expensive, particularly when dealing with very large data sets. In this paper, we present a computationally efficient approach for estimating the t-SVD by capitalizing on the correlations of the data in the temporal dimension. The approach proceeds by extending our prior work on fast eigenspace decompositions by transforming the tensor data from the spatial domain to the spectral domain in order to obtain reduced order harmonic tensor. The t-SVD can then be applied in the transform domain thereby significantly reducing the computational burden. Experimental results which are presented on the extended Yale-B, COIL-100, and MNIST data sets show the proposed method provides considerable computational savings with the approximated subspaces that are nearly the same as the true subspaces as computed via the t-SVD.","PeriodicalId":6750,"journal":{"name":"2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)","volume":"64 1","pages":"527-533"},"PeriodicalIF":0.0,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"81380157","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：