Researcher:
Akgün, Barış

Profile Picture
ORCID

Job Title

Faculty Member

First Name

Barış

Last Name

Akgün

Name

Name Variants

Akgün, Barış

Email Address

Birth Date

Filters

2018 - 201942020 - 20238

Advanced Search

Filter by

Settings

Researcher Of Publications: search.filters.isAuthorOfPublication.099c316d-ed14-4fd8-ab7d-a64702e6b796

Search Results

Now showing 1 - 10 of 12
  • Placeholder
    PublicationMetadata only
    Distributed deep reinforcement learning with wideband sensing for dynamic spectrum access
    (Ieee, 2020) Ucar, Seyhan; N/A; Department of Computer Engineering; Department of Electrical and Electronics Engineering; Kaytaz, Umuralp; Akgün, Barış; Ergen, Sinem Çöleri; PhD Student; Faculty Member; Faculty Member; Department of Computer Engineering; Department of Electrical and Electronics Engineering; Graduate School of Sciences and Engineering; College of Engineering; College of Engineering; N/A; 258784; 7211
    Wideband spectrum sensing (WBS) has been a critical issue for communication system designers and specialists to monitor and regulate the wireless spectrum. Detecting and identifying the existing signals in a continuous manner enable orchestrating signals through all controllable dimensions and enhancing resource usage efficiency. This paper presents an investigation on the application of deep learning (DL)-based algorithms within the WBS problem while also providing comparisons to the conventional recursive thresholding-based solution. For this purpose, two prominent object detectors, You Only Learn One Representation (YOLOR) and Detectron2, are implemented and fine-tuned to complete these tasks for WBS. The power spectral densities (PSDs) belonging to over-the-air (OTA) collected signals within the wide frequency range are recorded as images that constitute the signal signatures (i.e., the objects of interest) and are fed through the input of the above-mentioned learning and evaluation processes. The main signal types of interest are determined as the cellular and broadcast types (i.e., GSM, UMTS, LTE and Analogue TV) and the single-tone. With a limited amount of captured OTA data, the DL-based approaches YOLOR and Detectron2 are seen to achieve a classification rate of 100% and detection rates of 85% and 69%, respectively, for a nonzero intersection over union threshold. The preliminary results of our investigation clearly show that both object detectors are promising to take on the task of wideband signal detection and identification, especially after an extended data collection campaign.
  • Placeholder
    PublicationMetadata only
    Reward learning from very few demonstrations
    (Ieee-Inst Electrical Electronics Engineers Inc, 2021) N/A; Eteke, Cem; Kebüde, Doğancan; Akgün, Barış; PhD Student; Graduate School of Sciences and Engineering; N/A
    This article introduces a novel skill learning framework that learns rewards from very few demonstrations and uses them in policy search (PS) to improve the skill. The demonstrations are used to learn a parameterized policy to execute the skill and a goal model, as a hidden Markov model (HMM), to monitor executions. The rewards are learned from the HMM structure and its monitoring capability. The HMM is converted to a finite-horizon Markov reward process (MRP). A Monte Carlo approach is used to calculate its values. Then, the HMM and the values are merged into a partially observable MRP to obtain execution returns to be used with PS for improving the policy. In addition to reward learning, a black box PS method with an adaptive exploration strategy is adopted. The resulting framework is evaluated with five PS approaches and two skills in simulation. The results show that the learned dense rewards lead to better performance compared to sparse monitoring signals, and using an adaptive exploration lead to faster convergence with higher success rates and lower variance. The efficacy of the framework is validated in a real-robot settings by improving three skills to complete success from complete failure using learned rewards where sparse rewards failed completely.
  • Placeholder
    PublicationMetadata only
    Banking order classification and information extraction
    (Institute of Electrical and Electronics Engineers Inc., 2022) Bakır, Veli Oğuzalp; Çağatay, İlhan; Güven, Melih; Koras, Murat; Department of Industrial Engineering; Department of Computer Engineering; Gönen, Mehmet; Akgün, Barış; Faculty Member; Faculty Member; Department of Industrial Engineering; Department of Computer Engineering; College of Engineering; College of Engineering; 237468; 258784
    This study presents a system to classify banking orders from customers and to determine the transaction parameters of these orders using machine learning techniques. The presented system uses optical character recognition and shape detection technologies to extract texts and tables from images i.e., scanned email attachments and fax images. Then, in the classification phase, texts are vectorized with the TF-IDF approach after preprocessing and are classified using support vector machines. The orders classified as money transfer are sent to the information extraction module and the parameters of the transaction (sender information, recipient information, amount and description) are determined using named entity recognition methods. Finally, this information is sent directly to an operator's screen for her to check and confirm the parameters and execute the money transfer operation. This system is implemented in a medium-large scale bank in Turkey. This system, which yields high classification and information extraction performance, is expected to save a significant amount of workload for the bank, speed up the order execution process and increase customer satisfaction. The system is currently deployed and being validated online.
  • Placeholder
    PublicationMetadata only
    Chain FL: Decentralized federated machine learning via blockchain
    (Ieee, 2020) Masry, Ahmed; Department of Electrical and Electronics Engineering; Department of Computer Engineering; Department of Computer Engineering; Department of Computer Engineering; Korkmaz, Caner; Koçaş, Halil Eralp; Uysal, Ahmet; Özkasap, Öznur; Akgün, Barış; Undergraduate Student; Undergraduate Student; Undergraduate Student; Faculty Member; Faculty Member; Department of Electrical and Electronics Engineering; Department of Computer Engineering; College of Engineering; College of Engineering; College of Engineering; College of Engineering; College of Engineering; N/A; N/A; N/A; 113507; 258784
    Federated learning is a collaborative machine learning mechanism that allows multiple parties to develop a model without sharing the training data. It is a promising mechanism since it empowers collaboration in fields such as medicine and banking where data sharing is not favorable due to legal, technical, ethical, or safety issues without significantly sacrificing accuracy. In centralized federated learning, there is a single central server, and hence it has a single point of failure. Unlike centralized federated learning, decentralized federated learning does not depend on a single central server for the updates. In this paper, we propose a decentralized federated learning approach named Chain FL that makes use of the blockchain to delegate the responsibility of storing the model to the nodes on the network instead of a centralized server. Chain FL produced promising results on the MNIST digit recognition task with a maximum 0.20% accuracy decrease, and on the CIFAR-10 image classification task with a maximum of 2.57% accuracy decrease as compared to non-FL counterparts.
  • Placeholder
    PublicationMetadata only
    Distributed landmark placement in P2P networks
    (Institute of Electrical and Electronics Engineers (IEEE), 2018) N/A; Department of Computer Engineering; Department of Computer Engineering; Boshrooyeh, Sanaz Taheri; Özkasap, Öznur; Akgün, Barış; PhD Student; Faculty Member; Faculty Member; Department of Computer Engineering; Graduate School of Sciences and Engineering; College of Engineering; College of Engineering; N/A; 113507; 258784
    Peer-to-peer (P2P) paradigm is a promising way to provide services to the masses for a very low cost, and hence, P2P services have been gaining significant attention. Since P2P services usually operate over their users' resources, without using central servers, their performance is becoming of interest to researchers. Several existing solutions rely on supernodes, called landmarks, to enhance system performance. The landmarks are placed manually based on the density and the distribution of the nodes in the corresponding P2P network, and landmark locations are presumed to be determined before system setup. However, this assumption is not realistic since, in a P2P system, no global view about the peers' locations exists. Hence, the best landmark locations would not be known a priori. We propose a family of distributed landmark placement algorithms, called DLP, to address this issue. We implement four versions of the DLP family and evaluate their landmark placements in a simulation environment. We define two performance metrics to assess their performance as compared to manually placed landmarks. Our results show that the DLP algorithms can generate landmark locations that are on par with the manual placement and significantly reduce the landmark-to-peer latencies.
  • Placeholder
    PublicationMetadata only
    ATM allocation using decision tree-based algorithms
    (Ieee, 2021) Yurdakul, Hazal Hasret; Kasikci, Kerem; Cagatay, Ilhan; Guven, Melih; Koras, Murat; Department of Computer Engineering; Department of Industrial Engineering; Akgün, Barış; Gönen, Mehmet; Faculty Member; Faculty Member; Department of Computer Engineering; Department of Industrial Engineering; College of Engineering; College of Engineering; 258784; 237468
    Automated teller machines (ATM's) make it possible for customers to fulfill their financial operations easily and reduces the workload of bank branches if they are placed in convenient locations. Banks need to have ATMs allocated in favorable locations regarding customer concerns. In this study, the ATM allocation problem is handled using decision tree-based algorithms. To solve the problem, a machine learning algorithm should learn the characteristics of each defined region and understand factors affecting the business performance. Therefore, a grid system is designed by dividing Turkey by imaginary lines. Imaginary lines constitute small grids passing through each one-thousandth of a latitude degree and one-thousandth of a longitude degree. For each grid rectangle, the characteristics of the customers living or wandering there, the point of interest locations around the area, and the existence of the competitors' ATMs are determined. Then, algorithms are trained and scored using decision tree-based algorithms. To decide suitable grid areas for installment, the business value is calculated for each grid. A heat map presenting the scores of the whole country is created for visualization purposes. The proposed framework can be used to better allocate ATMs all around in Turkey./Öz: Bankamatikler (ATM), banka müşterilerinin finansal işlemlerini kolayca gerçekleştirebilmelerine imkan sunar, uygun noktalara yerleştirildiklerinde de şubelerin işyükünün azalmasına yardımcı olurlar. Müşteri istekleri göz önünde bulundurulduğunda bankalar ATM’lerini kolay erişilebilir konumlara yerleş- tirmelidir. Bu çalışmada ATM yerleşimi problemi karar ağacı temelli algoritmalarla işlenmiştir. Bu sorunu çözmek amacıyla yapay ögrenme algoritması tanımlanacak her bölgenin karak- teristiğini ve iş performansını etkileyen faktörleri öğrenmelidir. Bu sebeple, Türkiye’yi hayali çizgilerle ayıran bir ızgara sistemi tasarlandı. Hayali çizgiler bir enlem derecesinin ve bir boylam derecesinin binde birlik parçalarından geçen çizgilerden oluşan küçük ızgaralar meydana getirir. Her bir ızgara dikdörtgeni için, burada yaşayan veya gezen müşterilerin karakteristigi, o alan ve alanın çevresindeki ilgi alanları ve rakip bankaların ATM’lerinin varlığı belirlendi. Sonrasında, karar ağacı temelli algoritmalar kullanarak eğitildi ve değerlendirildi. Yerleşime uygun alanlara karar vermek amacıyla, her ızgaranın iş değeri hesaplandı. Görselleştirme maksatlı bütün ülkenin değerlerini gösteren bir ısı haritası oluşturuldu. Artık, önerilen sistem bütün Türkiye’de ATM’leri daha iyi yerleştirmek için kullanılabilir.
  • Placeholder
    PublicationMetadata only
    FLAGS framework for comparative analysis of federated learning algorithms
    (Elsevier, 2022) N/A; N/A; Department of Computer Engineering; Department of Computer Engineering; Lodhi, Ahnaf Hannan; Akgün, Barış; Özkasap, Öznur; PhD Student; Faculty Member; Faculty Member; Department of Computer Engineering; Graduate School of Sciences and Engineering; College of Engineering; College of Engineering; N/A; 258784; 113507
    Federated Learning (FL) has become a key choice for distributed machine learning. Initially focused on centralized aggregation, recent works in FL have emphasized greater decentralization to adapt to the highly heterogeneous network edge. Among these, Hierarchical, Device-to-Device and Gossip Federated Learning (HFL, D2DFL & GFL respectively) can be considered as foundational FL algorithms employing fundamental aggregation strategies. A number of FL algorithms were subsequently proposed employing multiple fundamental aggregation schemes jointly. Existing research, however, subjects the FL algorithms to varied conditions and gauges the performance of these algorithms mainly against Federated Averaging (FedAvg) only. This work consolidates the FL landscape and offers an objective analysis of the major FL algorithms through a comprehensive cross-evaluation for a wide range of operating conditions. In addition to the three foundational FL algorithms, this work also analyzes six derived algorithms. To enable a uniform assessment, a multi-FL framework named FLAGS: Federated Learning AlGorithms Simulation has been developed for rapid configuration of multiple FL algorithms. Our experiments indicate that fully decentralized FL algorithms achieve comparable accuracy under multiple operating conditions, including asynchronous aggregation and the presence of stragglers. Furthermore, decentralized FL can also operate in noisy environments and with a comparably higher local update rate. However, the impact of extremely skewed data distributions on decentralized FL is much more adverse than on centralized variants. The results indicate that it may not be necessary to restrict the devices to a single FL algorithm; rather, multi-FL nodes may operate with greater efficiency.
  • Placeholder
    PublicationMetadata only
    Communicative cues for reach-to-grasp motions: from humans to robots: robotics track
    (International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS), 2018) N/A; N/A; N/A; Department of Computer Engineering; Department of Computer Engineering; Kebüde, Doğancan; Eteke, Cem; Sezgin, Tevfik Metin; Akgün, Barış; Master Student; Master Student; Faculty Member; Faculty Member; Department of Computer Engineering; Graduate School of Sciences and Engineering; Graduate School of Sciences and Engineering; College of Engineering; College of Engineering; N/A; N/A; 18632; 258784
    Intent communication is an important challenge in the context of human-robot interaction. The aim of this work is to identify subtle non-verbal cues that make communication among humans fluent and use them to generate intent expressive robot motion. A human- human reach-to-grasp experiment (n = 14) identified two temporal and two spatial cues: (1) relative time to reach maximum hand aperture (AM), (2) overall motion duration (07), (3) exaggeration in motion (Exg), and (4) change in grasp modality (GM). Results showed there was statistically significant difference in the temporal cues between no-intention and intention conditions. In a follow-up experiment (n = 30), reach-to-grasp motions of a simulated robot containing different cue combinations were shown to the partici-pants. They were asked to guess the target object during robot's motion, based on the assumption that intent expressive motion would result in earlier and more accurate guesses. Results showed that, OT, GM and several cue combinations led to faster and more accurate guesses which imply they can be used to generate communicative motion. However, MA had no effect, and surprisingly Exg had a negative effect on expressiveness.
  • Placeholder
    PublicationMetadata only
    Corporate network analysis based on graph learning
    (Springer Science and Business Media Deutschland GmbH, 2023) Atan, E.; Duymaz, A.; Sarısözen, F.; Aydın, U.; Koraş, M.; Department of Computer Engineering; Department of Industrial Engineering; Akgün, Barış; Gönen, Mehmet; Faculty Member; Faculty Member; Department of Computer Engineering; Department of Industrial Engineering; College of Engineering; College of Engineering; 258784; 237468
    We constructed a financial network based on the relationships of the customers in our database with our other customers or other bank customers using our large-scale data set of money transactions. There are two main aims in this study. Our first aim is to identify the most profitable customers by prioritizing companies in terms of centrality based on the volume of money transfers between companies. This requires acquiring new customers, deepening existing customers and activating inactive customers. Our second aim is to determine the effect of customers on related customers as a result of the financial deterioration in this network. In this study, while creating the network, a data set was created over money transfers between companies. Here, text similarity algorithms were used while trying to match the company title in the database with the title during the transfer. For customers who are not customers of our bank, information such as IBAN numbers are assigned as unique identifiers. We showed that the average profitability of the top 30% customers in terms of centrality is five times higher than the remaining customers. Besides, the variables we created to examine the effect of financial disruptions on other customers contributed an additional 1% Gini coefficient to the model that the bank is currently using even if it is difficult to contribute to a strong model that already works with a high Gini coefficient. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.
  • Placeholder
    PublicationMetadata only
    Communicative cues for reach-to-grasp motions: From humans to robots
    (Assoc Computing Machinery, 2018) N/A; N/A; N/A; Department of Computer Engineering; Department of Computer Engineering; Kebüde, Doğancan; Eteke, Cem; Sezgin, Tevfik Metin; Akgün, Barış; Master Student; PhD Student; Faculty Member; Faculty Member; Department of Computer Engineering; Graduate School of Sciences and Engineering; Graduate School of Sciences and Engineering; College of Engineering; College of Engineering; N/A; N/A; 42946; 258784
    Intent communication is an important challenge in the context of human-robot interaction. The aim of this work is to identify subtle non-verbal cues that make communication among humans fluent and use them to generate intent expressive robot motion. A human human reach-to-grasp experiment (n = 14) identified two temporal and two spatial cues: (1) relative time to reach maximum hand aperture (MA), (2) overall motion duration (OT), (3) exaggeration in motion (Exg), and (4) change in grasp modality (GM). Results showed there was statistically significant difference in the temporal cues between no-intention and intention conditions. In a follow-up experiment (n = 30), reach-to-grasp motions of a simulated robot containing different cue combinations were shown to the participants. They were asked to guess the target object during robot's motion, based on the assumption that intent expressive motion would result in earlier and more accurate guesses. Results showed that, OT, GM and several cue combinations led to faster and more accurate guesses which imply they can be used to generate communicative motion. However, MA had no effect, and surprisingly Exg had a negative effect on expressiveness.