Yazar "Dalyan, Tugba" seçeneğine göre listele

Listeleniyor 1 - 7 / 7
  • Küçük Resim Yok
    Öğe
    A comparative study of neural machine translation models for Turkish language
    (Ios Press, 2022) Ozdemir, Ozgur; Akin, Emre Salih; Velioglu, Riza; Dalyan, Tugba
    Machine translation (MT) is an important challenge in the fields of Computational Linguistics. In this study, we conducted neural machine translation (NMT) experiments on two different architectures. First, Sequence to Sequence (Seq2Seq) architecture along with a variation that utilizes attention mechanism is performed on translation task. Second, an architecture that is fully based on the self-attention mechanism, namely Transformer, is employed to perform a comprehensive comparison. Besides, the contribution of employing Byte Pair Encoding (BPE) and Gumbel Softmax distributions are examined for both architectures. The experiments are conducted on two different datasets: TED Talks that is one of the popular benchmark datasets for NMT especially among morphologically rich languages like Turkish and WMT18 News dataset that is provided by The Third Conference on Machine Translation (WMT) for shared tasks on various aspects of machine translation. The evaluation of Turkish-to-English translations' results demonstrate that the Transformer model with combination of BPE and Gumbel Softmax achieved 22.4 BLEU score on TED Talks and 38.7 BLUE score on WMT18 News dataset. The empirical results support that using Gumbel Softmax distribution improves the quality of translations for both architectures.
  • Küçük Resim Yok
    Öğe
    A Fully Automated SPICE-Compatible Netlist Extraction From Image Using Deep Learning and Image Preprocessing Techniques
    (Ieee-Inst Electrical Electronics Engineers Inc, 2026) Peker, Omer Baran; Toker, Emre; Ocal, Dogukan; Dalyan, Tugba; Afacan, Engin; Gokdel, Yigit Daghan
    This paper presents an automated framework for generating SPICE compatible netlists from both printed and hand-drawn circuit diagrams. The system combines advanced image preprocessing, deep learning based object detection, and contour based node analysis to address challenges such as inconsistent drawing styles, illumination variations, and non-standardized symbols. A unified preprocessing module incorporating denoising, contrast enhancement, adaptive thresholding, morphological filtering, and skeletonization ensures robust inputs for downstream tasks. Multiple YOLO (You Only Look Once) architectures were trained and evaluated, demonstrating strong performance across subtasks: YOLOv8L achieved 97.50% for transistor detection, YOLOv11L reached 98.55% for terminal segmentation, YOLOv11X attained 96.13% for voltage segmentation, and YOLOv8L obtained 99.23% for ground detection. These results confirm the framework's reliability in symbol interpretation. Beyond component-level recognition, the system integrates a specialized transistor terminal segmentation model and an advanced contour-based node detection module, enabling the accurate extraction of connectivity, even in dense, multi-component circuits. A novel validation mechanism further enhances robustness by fully automated simulating generated netlists in LTSpice and comparing node voltages with those of reference designs. Experimental evaluation demonstrates superior performance on printed diagrams (93.33% accuracy) and competitive performance on hand-drawn sketches (85.33% accuracy), despite stylistic irregularities. Overall, the proposed pipeline provides a scalable and accurate end-to-end solution, reducing human error and ensuring functional equivalence. Its ability to process complex, large-scale hand-drawn schematics under diverse conditions highlights its contributions to Electronic Design Automation (EDA), industrial applications, and intelligent design assistance. In addition, the framework incorporates a fast and fully automated validation stage, where generated netlists are systematically simulated in LTspice and compared against reference designs. This ensures both structural correctness and functional equivalence, further enhancing robustness and reliability.
  • Küçük Resim Yok
    Öğe
    EMRES: A New EMotional RESpondent Robot
    (IEEE-Inst Electrical Electronics Engineers Inc, 2022) Sonmez, Elena Battini; Han, Hasan; Karadeniz, Oguzcan; Dalyan, Tugba; Sarioglu, Baykal
    The aim of this work is to design an artificial empathetic system and to implement it into an EMotional RESpondent (EMRES) robot, called EMRES. Rather than mimic the expression detected in the human partner, the proposed system achieves a coherent and consistent emotional trajectory resulting in a more credible human-agent interaction. Inspired by developmental robotics theory, EMRES has an internal state and a mood, which contribute in the evolution of the flow of emotions; at every episode, the next emotional state of the agent is affected by its internal state, mood, current emotion, and the expression read in the human partner. As a result, EMRES does not imitate, but it synchronizes to the emotion expressed by the human companion. The agent has been trained to recognize expressive faces of the FER2013 database and it is capable of achieving 78.3% performance with wild images. Our first prototype has been implemented into a robot, which has been created for this purpose. An empirical study run with university students judged in a positive way the newly proposed artificial empathetic system.
  • Küçük Resim Yok
    Öğe
    From Pixels to Spectra: Predicting Wine Colorimetric Characteristics Through Machine Learning Models
    (Springer, 2026) Erdemir, Naz; Sinop, Celal Deniz; Uysal, Reyhan Selin; Dalyan, Tugba
    This study employs digital image processing and machine learning techniques to predict all colorimetric characteristics (color intensity, density, tonality, and color index percentages) of colored wines in a novel, cost-effective, and rapid manner. To determine the values of colorimetric characteristics, ultraviolet-visible (UV-Vis) absorbance was measured at three key wavelengths (A420, A520, and A620) using UV-Vis spectrophotometry, corresponding to the yellow, red, and blue color percentages, respectively. Simultaneously, the pictures of 86 wine samples were acquired, and the corresponding RGB and HSV color values were extracted from the images to serve as input features for multiple regression models. The models developed included principal component regression, k-nearest neighbors, linear regression, decision tree, random forest, and partial least squares (PLS). Among the models, random forest outperformed PLS in predicting A620 absorbance value due to its ability to capture non-linear patterns, whereas PLS demonstrated greater accuracy (R2 > 0.95) in predicting the A420 and A520 absorbance values. According to feature selection, hue and saturation had the biggest impact on prediction accuracy. By determining absorbance values using the developed models, the complete colorimetric characteristics of the wine samples can be calculated, enabling the evaluation of their physicochemical parameters during the fermentation process or post-fermentation. As a result, all the models, improved, offer a promising alternative for quick, easy, and scalable prediction methods by reducing measurement time, eliminating the need for laboratory instruments, and introducing a new methodology to complement conventional spectroscopic techniques, with potential applications in consumer-level analysis and the process of wine quality control.
  • Küçük Resim Yok
    Öğe
    Interval-Valued Pythagorean Fuzzy AHP&TOPSIS for ERP Software Selection
    (Springer International Publishing Ag, 2022) Dalyan, Tugba; Otay, Irem; Gulada, Mehmet
    The selection of an Enterprise Resource Planning (ERP) system is considerably important since it has effects on the productivity of companies. This paper aims to choose the best decision for ERP software over many conflicting criteria by using Pythagorean fuzzy (PF) Analytic Hierarchy Process (AHP) that integrated with Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) method. The proposed model includes a hierarchical structure with four main criteria which are finance, technology, usability and corporate, twenty two sub-criteria and several alternatives. In this study, the evaluations are expressed using interval-valued PF sets that consist of membership and non-membership values, where the square sum of these degrees is at most 1. The weights of the criteria are computed using interval-valued PF AHP. Then, fuzzy TOPSIS method is utilized to evaluate alternatives by considering distances of alternatives to negative and positive ideal solutions (NIS, PIS), respectively. This attempt using interval-valued PF AHP-TOPSIS is deemed to be a significant contribution toward ERP selection problem.
  • Küçük Resim Yok
    Öğe
    Smart Solutions for Mega-Cities: Utilizing Long Short-Term Memory and Multi-Head Attention in Parking Prediction
    (Mdpi, 2024) Kemik, Hasan; Dalyan, Tugba; Aydogan, Murat
    Finding a parking space is a major concern in megacities, causing drivers to lose time and money while contributing to atmospheric pollution and global warming. This study proposes a method to predict parking slot availability by comparing Long Short-Term Memory (LSTM) and Multi-Head Attention (MHA) methods using the CityPulse Smart City Datasets. The initial experiments assessed the impact of pollution and time features on prediction accuracy. In a subsequent experiment, the dataset was expanded by incorporating weather-related features and a broader time range while excluding pollution and time features, as informed by the initial results. Various experiments were conducted with different parameters, such as model depth and activation functions. The results demonstrated that MHA outperformed LSTM in predicting occupancy rates, achieving a Mean Absolute Error (MAE) score of 0.0589 on the extended dataset. This study marks a pioneering effort in using MHA for real-time parking occupancy prediction, showcasing significant success with fewer parameters and a smaller model size.
  • Küçük Resim Yok
    Öğe
    Unified benchmark for zero-shot Turkish text classification
    (Elsevier Sci Ltd, 2023) celik, Emrecan; Dalyan, Tugba
    Effective learning schemes such as fine-tuning, zero-shot, and few-shot learning, have been widely used to obtain considerable performance with only a handful of annotated training data. In this paper, we presented a unified benchmark to facilitate the problem of zeroshot text classification in Turkish. For this purpose, we evaluated three methods, namely, Natural Language Inference, Next Sentence Prediction and our proposed model that is based on Masked Language Modeling and pre-trained word embeddings on nine Turkish datasets for three main categories: topic, sentiment, and emotion. We used pre-trained Turkish monolingual and multilingual transformer models which can be listed as BERT, ConvBERT, DistilBERT and mBERT. The results showed that ConvBERT with the NLI method yields the best results with 79% and outperforms previously used multilingual XLM-RoBERTa model by 19.6%. The study contributes to the literature using different and unattempted transformer models for Turkish and showing improvement of zero-shot text classification performance for monolingual models over multilingual models.