Betül Ay
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Betül Ay has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Galip Aydin
Department of Computer Engineering, Faculty of Engineering, Fırat University, 23119 Elazığ, Türkiye
5 shared publications Galip Aydin
Department of Computer Engineering, Firat University, Elazig, Turkey
3 shared publications- Ibrahim Riza Hallac
Alanya Alaaddin Keykubat Üniversitesi
2 shared publications - Mehmet Ata Öztürk
Department of Physical Education and Sports, Middle East Technical University, 06800 Ankara, Türkiye
2 shared publications Beyda Tasar
Department of Mechatronics Engineering, Firat University, Elazig, Turkey
1 shared publication- Cihan Turker
Department of Otolaryngology, Private Medical Hospital, Elazig 23040, Turkey
1 shared publication
Latest Publications
- Open Access
Skeleton-Based Activity Recognition for Children with Autism Using Graph Convolutional Networks
Movement-based and physical activity programs are central tools in autism intervention, so recognizing the activities a child performs during therapy is valuable for objective progress tracking. Manual monitoring of these sessions is time-consuming and subjective, and raw videos raise privacy concerns because it shows identifiable children. We address autism therapeutic activity recognition from privacy-preserving 2D skeletons, and we focus on the practical difficulty of how several therapeutic activities differ only in subtle motion details. As a backbone, we adopt ProtoGCN, a graph convolutional network that represents each action as a combination of learnable motion prototypes. However, this contrastive backbone organizes all classes at once, so it does not enforce a margin between the few pairs that remain entangled after training. We therefore introduce a Refine–Confusable (RC) module, a training-only regularizer that pushes apart the empirically most-confused class pairs using a hinge-margin loss over momentum-updated class centroids. The module changes neither the backbone nor the inference cost. On the MMASD dataset, restricted to the ten-class 2D-skeleton configuration, the RC module improves the base model across random, session-independent, and subject-independent evaluation. The gain is largest on the strictest subject-independent split and a clip-level analysis confirms that this improvement is statistically significant. Under the protocol-matched holdout, the method reaches 96.30% accuracy with 0.959 macro-F1, surpassing recent 2D-skeleton baselines while keeping a lightweight and privacy-preserving modality. The improvements are modest, as expected on a small clinical dataset, and t-SNE and prototype visualizations show that the learned representation is discriminative and interpretable.
Open AccessSkeleton-Based Activity Recognition for Children with Autism Using Graph Convolutional Networks
Movement-based and physical activity programs are central tools in autism intervention, so recognizing the activities a child performs during therapy is valuable for objective progress tracking. Manual monitoring of these sessions is time-consuming and subjective, and raw videos raise privacy concerns because it shows identifiable children. We address autism therapeutic activity recognition from privacy-preserving 2D skeletons, and we focus on the practical difficulty of how several therapeutic activities differ only in subtle motion details. As a backbone, we adopt ProtoGCN, a graph convolutional network that represents each action as a combination of learnable motion prototypes. However, this contrastive backbone organizes all classes at once, so it does not enforce a margin between the few pairs that remain entangled after training. We therefore introduce a Refine–Confusable (RC) module, a training-only regularizer that pushes apart the empirically most-confused class pairs using a hinge-margin loss over momentum-updated class centroids. The module changes neither the backbone nor the inference cost. On the MMASD dataset, restricted to the ten-class 2D-skeleton configuration, the RC module improves the base model across random, session-independent, and subject-independent evaluation. The gain is largest on the strictest subject-independent split and a clip-level analysis confirms that this improvement is statistically significant. Under the protocol-matched holdout, the method reaches 96.30% accuracy with 0.959 macro-F1, surpassing recent 2D-skeleton baselines while keeping a lightweight and privacy-preserving modality. The improvements are modest, as expected on a small clinical dataset, and t-SNE and prototype visualizations show that the learned representation is discriminative and interpretable.
- Open Access
- 24 June 2026
- Published by MDPI AG in Preprints.org
Skeleton-Based Activity Recognition for Children with Autism Using Graph Convolutional Networks
Movement-based and physical activity programs are central tools in autism intervention, so recognizing the activities a child performs during therapy is valuable for objective progress tracking. Manual monitoring of these sessions is time-consuming and subjective, and raw video raises privacy concerns because it shows identifiable children. We address autism therapy activity recognition from privacy-preserving 2D skeletons, and we focus on the practical difficulty that several therapy activities differ only in subtle motion details. As a backbone we adopt ProtoGCN, a graph convolutional network that represents each action as a combination of learnable motion prototypes. This contrastive backbone organizes all classes at once, however, so it does not enforce a margin between the few pairs that remain entangled after training. We therefore introduce a Refine-Confusable (RC) module, a training-only regularizer that pushes apart the empirically most-confused class pairs using a hinge-margin loss over momentum-updated class centroids. The module changes neither the backbone nor the inference cost. On the MMASD dataset restricted to the ten-class 2D-skeleton configuration, the RC module improves the base model under both splitting regimes, with a gain of 1.21 points of pooled accuracy on the stricter subject-independent split. Under the protocol-matched holdout it reaches 96.30% accuracy with 0.959 macro-F1, which surpasses recent 2D-skeleton baselines while it keeps a lightweight and privacy-preserving modality. A per-fold analysis shows that this gain is consistent across folds rather than tied to a single split, and t-SNE and prototype visualizations show that the learned representation is discriminative and interpretable.
- 24 June 2026
- Published by MDPI AG in Preprints.org
Open AccessSkeleton-Based Activity Recognition for Children with Autism Using Graph Convolutional Networks
Movement-based and physical activity programs are central tools in autism intervention, so recognizing the activities a child performs during therapy is valuable for objective progress tracking. Manual monitoring of these sessions is time-consuming and subjective, and raw video raises privacy concerns because it shows identifiable children. We address autism therapy activity recognition from privacy-preserving 2D skeletons, and we focus on the practical difficulty that several therapy activities differ only in subtle motion details. As a backbone we adopt ProtoGCN, a graph convolutional network that represents each action as a combination of learnable motion prototypes. This contrastive backbone organizes all classes at once, however, so it does not enforce a margin between the few pairs that remain entangled after training. We therefore introduce a Refine-Confusable (RC) module, a training-only regularizer that pushes apart the empirically most-confused class pairs using a hinge-margin loss over momentum-updated class centroids. The module changes neither the backbone nor the inference cost. On the MMASD dataset restricted to the ten-class 2D-skeleton configuration, the RC module improves the base model under both splitting regimes, with a gain of 1.21 points of pooled accuracy on the stricter subject-independent split. Under the protocol-matched holdout it reaches 96.30% accuracy with 0.959 macro-F1, which surpasses recent 2D-skeleton baselines while it keeps a lightweight and privacy-preserving modality. A per-fold analysis shows that this gain is consistent across folds rather than tied to a single split, and t-SNE and prototype visualizations show that the learned representation is discriminative and interpretable.
- Open Access
- 29 May 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (11) , 5435
Benchmarking Ollama and vLLM for Concurrent LLM Serving: A Multi-Scenario Evaluation of Performance and Scalability
Serving LLMs to many concurrent users gives rise to significant challenges for latency, throughput, and scalability. This paper presents a systematic and reproducible benchmark of two widely used large language model (LLM) serving frameworks, Ollama and vLLM, under concurrent workloads. We conducted a controlled comparison by running both frameworks on a single NVIDIA H100 80 GB GPU with the same model weights (Qwen3-4B) and inference configurations. We then evaluated them using five open benchmark datasets across four scenarios consisting of baseline question answering, complex reasoning, streaming interaction, and stress testing. Each scenario was executed under increasing concurrency levels, ranging from light loads to high-concurrency stress, to measure end-to-end latency, throughput, time-to-first-token (TTFT), success rate, and resource usage. Our experiments show that vLLM clearly outperforms Ollama across all four scenarios, achieving a 100% request success rate. It delivers 20–29 times higher throughput and 8–19 times lower P95 latency, and it completes every request successfully. Additionally, vLLM produced the first token within 0.5–3.5 s and remained stable at up to 100 concurrent users. Ollama, by contrast, required 54–122 s for the first token, hit a concurrency bottleneck near 10 users, and exhibited a timeout-based error rate of 13–30.06% under heavier loads. Notably, both frameworks demonstrated significant memory growth during extended endurance tests, necessitating careful monitoring in long-running deployments. Overall, for a single model on a single GPU, vLLM is highly suitable for high-concurrency serving, while Ollama remains a practical choice for lightweight, local, or developmental workflows.
- 29 May 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (11) , 5435
Open AccessBenchmarking Ollama and vLLM for Concurrent LLM Serving: A Multi-Scenario Evaluation of Performance and Scalability
Serving LLMs to many concurrent users gives rise to significant challenges for latency, throughput, and scalability. This paper presents a systematic and reproducible benchmark of two widely used large language model (LLM) serving frameworks, Ollama and vLLM, under concurrent workloads. We conducted a controlled comparison by running both frameworks on a single NVIDIA H100 80 GB GPU with the same model weights (Qwen3-4B) and inference configurations. We then evaluated them using five open benchmark datasets across four scenarios consisting of baseline question answering, complex reasoning, streaming interaction, and stress testing. Each scenario was executed under increasing concurrency levels, ranging from light loads to high-concurrency stress, to measure end-to-end latency, throughput, time-to-first-token (TTFT), success rate, and resource usage. Our experiments show that vLLM clearly outperforms Ollama across all four scenarios, achieving a 100% request success rate. It delivers 20–29 times higher throughput and 8–19 times lower P95 latency, and it completes every request successfully. Additionally, vLLM produced the first token within 0.5–3.5 s and remained stable at up to 100 concurrent users. Ollama, by contrast, required 54–122 s for the first token, hit a concurrency bottleneck near 10 users, and exhibited a timeout-based error rate of 13–30.06% under heavier loads. Notably, both frameworks demonstrated significant memory growth during extended endurance tests, necessitating careful monitoring in long-running deployments. Overall, for a single model on a single GPU, vLLM is highly suitable for high-concurrency serving, while Ollama remains a practical choice for lightweight, local, or developmental workflows.
- 1 January 2026
- Published by SETSCI in International Journal of Multidisciplinary Studies and Innovative Technologies
- Vol. 10 (1) , 72-78
Ağ Trafiği Anomali Tespitinde MLOps Tabanlı Makine Öğrenmesi Yaklaşımlarının Karşılaştırmalı Analizi
Ağ trafiği anomali tespiti, siber güvenlik sistemlerinde kritik bir rol oynamaktadır. Bu çalışmada ağ trafiği anomali tespiti problemi için farklı makine öğrenmesi modüllerinin performansı, uçtan uca bir MLOps platformu üzerinde karşılaştırmalı olarak değerlendirilmiştir. Çalışma kapsamında TensorFlow Decision Forests (TFDF), LightGBM, System AutoML ve H2O AutoML modülleri aynı veri seti ve deney protokolü kullanılarak analiz edilmiştir. Modellerin performansı Accuracy, AUC, Precision, Recall, F1, Kappa ve MCC metrikleri ile değerlendirilmiştir. Deneysel sonuçlar, TFDF Gradient Boosted Trees modelinin %92.50 doğruluk skoru ile en dengeli performansı sağladığını göstermektedir. Çalışma, model seçiminin yalnızca doğruluk metriğine göre yapılmasının yeterli olmadığını ve özellikle güvenlik uygulamalarında Recall ve F1 dengesinin kritik önem taşıdığını ortaya koymaktadır. Ayrıca MLOps platformu üzerinde çalışan farklı modüllerin deneylerin tekrarlanabilirliğini ve karşılaştırılabilirliğini artırdığı gösterilmiştir.
- 1 January 2026
- Published by SETSCI in International Journal of Multidisciplinary Studies and Innovative Technologies
- Vol. 10 (1) , 72-78
Ağ Trafiği Anomali Tespitinde MLOps Tabanlı Makine Öğrenmesi Yaklaşımlarının Karşılaştırmalı Analizi
Ağ trafiği anomali tespiti, siber güvenlik sistemlerinde kritik bir rol oynamaktadır. Bu çalışmada ağ trafiği anomali tespiti problemi için farklı makine öğrenmesi modüllerinin performansı, uçtan uca bir MLOps platformu üzerinde karşılaştırmalı olarak değerlendirilmiştir. Çalışma kapsamında TensorFlow Decision Forests (TFDF), LightGBM, System AutoML ve H2O AutoML modülleri aynı veri seti ve deney protokolü kullanılarak analiz edilmiştir. Modellerin performansı Accuracy, AUC, Precision, Recall, F1, Kappa ve MCC metrikleri ile değerlendirilmiştir. Deneysel sonuçlar, TFDF Gradient Boosted Trees modelinin %92.50 doğruluk skoru ile en dengeli performansı sağladığını göstermektedir. Çalışma, model seçiminin yalnızca doğruluk metriğine göre yapılmasının yeterli olmadığını ve özellikle güvenlik uygulamalarında Recall ve F1 dengesinin kritik önem taşıdığını ortaya koymaktadır. Ayrıca MLOps platformu üzerinde çalışan farklı modüllerin deneylerin tekrarlanabilirliğini ve karşılaştırılabilirliğini artırdığı gösterilmiştir.
- 30 September 2022
- Published by Firat Universitesi in Turkish Journal of Science and Technology
- Vol. 17 (2) , 435-450
Active Face Spoof Detection Using Image Distortion Analysis
With the rising use of facial recognition systems in a range of real-world scenarios and applications, attackers are also increasing their efforts, with a number of spoofing techniques emerging. As a result, developing a reliable spoof detection mechanism is critical. Active-based techniques have been shown to be good at finding spoofs, but they have a number of problems, such as being intrusive, expensive, hard to compute, not being able to be used in many situations, and usually needing extra hardware. This research presented an active-based robust spoof detection technique capable of detecting a wide range of media or 2D attacks while being less intrusive, less expensive, low in complexity, and more generalizable than other active-based techniques. It doesn't require any additional hardware, so it can easily be integrated into current systems. The distortion variations of video frames of the user's face collected at varying distances from the camera are analyzed to detect spoofing. Both the legitimate and spoof attack datasets were created using real-world facial photo and video data. The proposed approach achieved a spoof detection accuracy of 98.18% using both machine learning classifiers and a deep learning model, with an equal error rate and a half total error rate as low as 0.023 and 0.021, respectively.
- 30 September 2022
- Published by Firat Universitesi in Turkish Journal of Science and Technology
- Vol. 17 (2) , 435-450
Active Face Spoof Detection Using Image Distortion Analysis
With the rising use of facial recognition systems in a range of real-world scenarios and applications, attackers are also increasing their efforts, with a number of spoofing techniques emerging. As a result, developing a reliable spoof detection mechanism is critical. Active-based techniques have been shown to be good at finding spoofs, but they have a number of problems, such as being intrusive, expensive, hard to compute, not being able to be used in many situations, and usually needing extra hardware. This research presented an active-based robust spoof detection technique capable of detecting a wide range of media or 2D attacks while being less intrusive, less expensive, low in complexity, and more generalizable than other active-based techniques. It doesn't require any additional hardware, so it can easily be integrated into current systems. The distortion variations of video frames of the user's face collected at varying distances from the camera are analyzed to detect spoofing. Both the legitimate and spoof attack datasets were created using real-world facial photo and video data. The proposed approach achieved a spoof detection accuracy of 98.18% using both machine learning classifiers and a deep learning model, with an equal error rate and a half total error rate as low as 0.023 and 0.021, respectively.
- 13 June 2022
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 147, 105725
Automated classification of nasal polyps in endoscopy video-frames using handcrafted and CNN features
- 13 June 2022
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 147, 105725
Automated classification of nasal polyps in endoscopy video-frames using handcrafted and CNN features
- 4 May 2022
- Published by Springer Nature in Neural Computing and Applications
- Vol. 34 (18) , 16157-16168
Deep transfer learning-based visual classification of pressure injuries stages
- 4 May 2022
- Published by Springer Nature in Neural Computing and Applications
- Vol. 34 (18) , 16157-16168
Deep transfer learning-based visual classification of pressure injuries stages
- Open Access
- 2 March 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 10, 25114-25124
Open-Set Learning-Based Hologram Verification System Using Generative Adversarial Networks
In this study, we address the hologram authenticity challenge by introducing a novel deep-learning based end-to-end hologram verification system. The system ultimately makes the decision whether the hologram image captured from a mobile application is fake or not by employing a robust classifier. We built the system by training three major deep networks; generative networks, convolutional networks and region-based convolutional networks. One major challenge in this study was the lack of negative class samples or so-called fake holograms. To the best of our knowledge there are no publicly available fake hologram datasets and it is not clear how the attackers imitate the real holograms. Therefore, the negative class in the practical hologram classification task is actually “unknown” class, as it is unknown how to imitate holograms by attackers. We hereby consider the problem of hologram classification as in a similar logic to open-set recognition. To make hologram classifier more sensitive to forgery, we generate synthetic images using generative adversarial networks (GANs) to represent negative class. We conduct extensive and comparative experiments on the closed-set and open-set using the-state-of-the-art backbone convolutional neural networks (CNNs). The proposed system gives an impressive accuracy 97.5% and 79% for closed-set and open-set samples, respectively. The reported results show the strong generalization performance of the system for unknown samples.
- 2 March 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 10, 25114-25124
Open AccessOpen-Set Learning-Based Hologram Verification System Using Generative Adversarial Networks
In this study, we address the hologram authenticity challenge by introducing a novel deep-learning based end-to-end hologram verification system. The system ultimately makes the decision whether the hologram image captured from a mobile application is fake or not by employing a robust classifier. We built the system by training three major deep networks; generative networks, convolutional networks and region-based convolutional networks. One major challenge in this study was the lack of negative class samples or so-called fake holograms. To the best of our knowledge there are no publicly available fake hologram datasets and it is not clear how the attackers imitate the real holograms. Therefore, the negative class in the practical hologram classification task is actually “unknown” class, as it is unknown how to imitate holograms by attackers. We hereby consider the problem of hologram classification as in a similar logic to open-set recognition. To make hologram classifier more sensitive to forgery, we generate synthetic images using generative adversarial networks (GANs) to represent negative class. We conduct extensive and comparative experiments on the closed-set and open-set using the-state-of-the-art backbone convolutional neural networks (CNNs). The proposed system gives an impressive accuracy 97.5% and 79% for closed-set and open-set samples, respectively. The reported results show the strong generalization performance of the system for unknown samples.
- Open Access
- 13 June 2021
- Published by MDPI AG in Applied Sciences
- Vol. 11 (12) , 5489
User Representation Learning for Social Networks: An Empirical Study
Gathering useful insights from social media data has gained great interest over the recent years. User representation can be a key task in mining publicly available user-generated rich content offered by the social media platforms. The way to automatically create meaningful observations about users of a social network is to obtain real-valued vectors for the users with user embedding representation learning models. In this study, we presented one of the most comprehensive studies in the literature in terms of learning high-quality social media user representations by leveraging state-of-the-art text representation approaches. We proposed a novel doc2vec-based representation method, which can encode both textual and non-textual information of a social media user into a low dimensional vector. In addition, various experiments were performed for investigating the performance of text representation techniques and concepts including word2vec, doc2vec, Glove, NumberBatch, FastText, BERT, ELMO, and TF-IDF. We also shared a new social media dataset comprising data from 500 manually selected Twitter users of five predefined groups. The dataset contains different activity data such as comment, retweet, like, location, as well as the actual tweets composed by the users.
- 13 June 2021
- Published by MDPI AG in Applied Sciences
- Vol. 11 (12) , 5489
Open AccessUser Representation Learning for Social Networks: An Empirical Study
Gathering useful insights from social media data has gained great interest over the recent years. User representation can be a key task in mining publicly available user-generated rich content offered by the social media platforms. The way to automatically create meaningful observations about users of a social network is to obtain real-valued vectors for the users with user embedding representation learning models. In this study, we presented one of the most comprehensive studies in the literature in terms of learning high-quality social media user representations by leveraging state-of-the-art text representation approaches. We proposed a novel doc2vec-based representation method, which can encode both textual and non-textual information of a social media user into a low dimensional vector. In addition, various experiments were performed for investigating the performance of text representation techniques and concepts including word2vec, doc2vec, Glove, NumberBatch, FastText, BERT, ELMO, and TF-IDF. We also shared a new social media dataset comprising data from 500 manually selected Twitter users of five predefined groups. The dataset contains different activity data such as comment, retweet, like, location, as well as the actual tweets composed by the users.
- 26 July 2019
- Published by SETSCI
- Vol. 4(5), 7-10
Real-time Pedestrian Attribute Detection from Surveillance Cameras
In recent years, computer vision has taken great strides in understanding and recognition the visual scenes together with deep learning technologies. Object recognition is one of the key areas of the computer vision applications. It is mainly concerned with recognition and localization of specific objects in an image. There are open-source and pre-trained models for the detection of general objects (such as cars, persons, cats, dogs). However, it is necessary to develop problem-based algorithms and training with deep neural networks by creating annotated training data for the recognition of special objects (such as headscarf). Furthermore, objects come in different shapes, sizes, angles, colors and additional noise from changes in the realworld environment, perspective, lighting and shadows. Taking into account these problems and needs in the real world, this paper focuses on the development of problem-based deep neural networks algorithms and the creation of labeled and reliable training datasets for the objects to be recognized. The contribution of the paper is to make use of transfer learning with the optimized R-FCN and Faster R-CNN pre-trained models in order to recognize pedestrian attributes including hat, headscarf, eyeglasses, bag objects and gender from security cameras. The proposed detection model has been trained on large-scale labeled dataset using TensorFlow open source platform. The performance of the neural network model has been evaluated using Average Precision (AP) values for each class and over 75% Mean Average Precision (mAP) for all classes is achieved.
- 26 July 2019
- Published by SETSCI
- Vol. 4(5), 7-10
Real-time Pedestrian Attribute Detection from Surveillance Cameras
In recent years, computer vision has taken great strides in understanding and recognition the visual scenes together with deep learning technologies. Object recognition is one of the key areas of the computer vision applications. It is mainly concerned with recognition and localization of specific objects in an image. There are open-source and pre-trained models for the detection of general objects (such as cars, persons, cats, dogs). However, it is necessary to develop problem-based algorithms and training with deep neural networks by creating annotated training data for the recognition of special objects (such as headscarf). Furthermore, objects come in different shapes, sizes, angles, colors and additional noise from changes in the realworld environment, perspective, lighting and shadows. Taking into account these problems and needs in the real world, this paper focuses on the development of problem-based deep neural networks algorithms and the creation of labeled and reliable training datasets for the objects to be recognized. The contribution of the paper is to make use of transfer learning with the optimized R-FCN and Faster R-CNN pre-trained models in order to recognize pedestrian attributes including hat, headscarf, eyeglasses, bag objects and gender from security cameras. The proposed detection model has been trained on large-scale labeled dataset using TensorFlow open source platform. The performance of the neural network model has been evaluated using Average Precision (AP) values for each class and over 75% Mean Average Precision (mAP) for all classes is achieved.