Accurate and trustworthy diagnosis from chest radiographs remains a critical challenge in medical imaging. In this work, we propose a robust deep learning framework based on the Conformer architecture to address multi-label classification of thoracic diseases using the CheXpert dataset. The Conformer, which integrates convolutional and self-attention mechanisms, is particularly well-suited for modeling both local and global dependencies in high-resolution medical images. Our approach incorporates three key pillars for clinical readiness: (i) robustness, evaluated through adversarial perturbations and sensitivity analysis to ensure resilience to input noise; (ii) uncertainty quantification, enabling calibrated confidence estimates essential for high-stakes decision-making; and (iii) explainability, achieved through feature attribution techniques. Extensive experiments demonstrate that our model not only achieves competitive performance across multiple pathologies but also maintains stable predictions under distributional shifts and perturbations. These results highlight the potential of hybrid transformer-based models as reliable tools for real-world medical imaging applications, bridging the gap between performance and deployability.

Online hate speech has been linked to a global rise in violence against minorities, including incidents such as mass shootings, lynchings, and ethnic cleansing. Societies grappling with this issue, particularly when hate speech targets specific groups based on religion, race, ethnicity, culture, nationality, or migration status, face the challenge of balancing freedom of expression with the need for effective content moderation on widely used online platforms. In response to this challenge, we introduce a comprehensive hate speech dataset covering five distinct topics in Turkish: refugees, the Israel-Palestine conflict, anti-Greek sentiment in Turkey, ethnic or religious communities (Alevis, Armenians, Arabs, Jews, and Kurds), and LGBTI+, alongside one topic in Arabic (refugees). In addition, we develop state-of-the-art BERT-based models to address multiple dimensions of hate speech analysis, including hate category classification, hate intensity prediction, target identification, and hate speech span detection, enabling a comprehensive understanding of hateful content in online discourse.

Skin lesion classification is a critical medical image analysis task in which early diagnosis significantly improves survival rates. In this work, we address the parameter variability problem of Vision Transformer based models arising from small batch sizes at the end of training. We propose maintaining an auxiliary model that tracks the Exponential Moving Average (EMA) of the online model parameters during training. The proposed approach smooths parameter fluctuations and improves training stability without introducing any additional computational overhead. Experiments on the ISIC 2019 dataset demonstrate that, in addition to reducing parameter variability, the EMA model consistently achieves higher Macro-F1 scores (6.93%) and accuracy rates (2.67%) compared to the online model.

We present Variational Self-Supervised Learning (VSSL), a novel framework that combines variational inference with self-supervised learning to enable efficient, decoder-free representation learning. Unlike traditional VAEs that rely on input reconstruction via a decoder, VSSL symmetrically couples two encoders with Gaussian outputs. A momentum-updated teacher network defines a dynamic, data-dependent prior, while the student encoder produces an approximate posterior from augmented views. The reconstruction term in the ELBO is replaced with a cross-view denoising objective, preserving the analytical tractability of Gaussian KL divergence. We further introduce cosine-based formulations of KL and log-likelihood terms to enhance semantic alignment in high-dimensional latent spaces. Experiments on CIFAR-10, CIFAR-100, and ImageNet-100 show that VSSL achieves competitive or superior performance to leading self-supervised methods, including BYOL and MoCo V3. VSSL offers a scalable, probabilistically grounded approach to learning transferable representations without generative reconstruction, bridging the gap between variational modeling and modern self-supervised techniques.

Hate speech detection is a crucial task, especially on social media, where harmful content can spread quickly. Implementing machine learning models to automatically identify and address hate speech is essential for mitigating its impact and preventing its proliferation. The first step in developing an effective hate speech detection model is to acquire a high-quality dataset for training. Labeled data is essential for most natural language processing tasks, but categorizing hate speech is difficult due to the diverse and often subjective nature of hate speech, which can lead to varying interpretations and disagreements among annotators. This paper examines strategies for addressing annotator disagreement, an issue that has been largely overlooked. In particular, we evaluate various automatic approaches for aggregating multiple annotations, in the context of hate speech classification in Turkish tweets. Our work highlights the importance of the problem and provides state-of-the-art benchmark results for the detection and understanding of hate speech in online discourse.

We introduce a new approach in distributed learning, building on Hinton’s Forward-Forward (FF) algorithm to speed up the training of neural networks in distributed environments without losing accuracy. Unlike traditional methods that rely on forward and backward passes, the FF algorithm employs a dual forward pass strategy, eliminating the dependency among layers required during the backpropagation period, which prevents efficient parallelization of the training process. Although the original FF algorithm focused on its ability to match the performance of the backpropagation algorithm, this work aims to reduce the training time with pipeline parallelism. We propose three novel pipelined FF algorithms that speed up training 3.75 times on the MNIST dataset while maintaining accuracy when training a four-layer network with four compute nodes. These results show that FF is highly parallelizable and its potential in large-scale distributed/federated systems to enable faster training for larger and more complex models.

Vision Transformers (ViTs) have achieved great performance on a variety of computer vision tasks, including medical image classification. Fine-tuning these models is computationally costly. However, Parameter-Efficient Fine-Tuning (PEFT) methods address this limitation by fine-tuning a minimal number of parameters. Despite their efficiency, the success of these methods depends greatly on the hyperparameter selection. In this paper, we propose an automatic hyperparameter search method for BEFT using Optuna.To validate the effectiveness and generality of our approach, we applied this Auto-Optimized PEFT strategy to three state-of-the-art pretrained Vision Transformers: ViT-Base, ViT-Medium, and Swin Transformer. After optimization using our method, ViT-Base achieved an accuracy improvement of +10.58% (87.17%), ViT-Medium achieved an accuracy improvement of +7.09% (83.23%), and Swin Transformer achieved an accuracy improvement of +18.62% (85.57%). The results highlight the effectiveness of PEFT methods in combination with automatic hyperparameter tuning to successfully fine-tune large vision models even in resource-scarce medical environments.

This paper addresses the problem of automatic classification of dermoscopic skin lesion images into one of eight diagnoses that include multiple cancer types. The goal of this research was to evaluate methods to address problems of class imbalances and image variation in skin lesion classification, as well as to propose an ensemble model that performs well on multi-class metrics. The widely used dataset developed for the International Skin Imaging Collaboration (ISIC) challenge, ISIC 2019, was chosen to test the methods proposed in this paper. Methods evaluated in this paper include modern architectures such as Vision Transformer, Convolutional Neural Networks, and Vision Mamba, as well as training strategies such as data augmentation, test-time augmentation, and specialized loss functions addressing class imbalance. The best performance was obtained using an ensemble approach, achieving an overall accuracy of 74.80% and a macro F1 score of 74.30%. These results compare favorably to the state-of-the-art, where drops in performance with multi-class metrics have typically been observed.

Skin lesion classification is a key predictor for early skin cancer detection. But bringing deep learning models to resource-constrained environments, such as mobile or embedded systems, requires efficiency in both accuracy and memory utilization. We present, in this paper, a Convolutional Neural Network (CNN) with 8,550,983 parameters for skin lesion classification and a K-Means-based clustered quantization technique for compressing the model while maintaining high classification efficiency.Aside from quantizing schemes standard to the industry—8-bit int, 16-bit float, and 32-bit float—we apply K-Means quantization, where the weights of a given layer are clustered into summary centroids. This will cause lower precision and redundancy, but creates an efficient but small model representation.The K-Means clustering model still had good performance with an accuracy value of 97%, precision 97.20%, recall 97.09%, and F1-score 96.95%, and compression ratio 0.1. These results indicate that K-Means-based clustered quantization has a good solution for model size minimization without performance loss, making it a great candidate for real-time skin lesion classification on embedded or mobile systems. Index Terms—Convolutional Neural Network, Random Oversampling, K-means Quantization, Image classification, Skin lesion

We address the problem of retrieving relevant images from the web in response to a text query. Text-based image retrieval is a challenging multimodal problem that requires understanding of the user intent from a few keywords and a semantic understanding of the images to predict relevance. In this paper, we present the hybrid search approach of the Turkcell-Yaani Search Engine, which was the leading search engine for queries in the Turkish language in 2017-2022. With over 1 million monthly users and processing millions of requests per month, its hybrid image search approach demonstrates the practical relevance and effectiveness of integrating text and visual information for enhanced image retrieval. The initial stage of the system focuses on retrieving images by comparing the query text with textual information associated with previously crawled and indexed images. The second stage aims to address the limitations of relying solely on text-based information by incorporating the post-processing steps based on semantic analysis of retrieved images. This stage, which is the main contribution of this paper, involves clustering, filtering and reordering the retrieved images using their embedded representations that are obtained from pretrained deep neural networks. The experimental results show that incorporating visual content improves the performance of the text-based retrieval engine significantly.

Signatures have long been used as a basic means of authentication. With the increase in digitization and document forgery, signature verification has gained greater importance than ever before. This study introduces the SSDV (Sabanci Signature Detection and Verification) dataset, developed for signature detection and verification. SSDV provides a rich dataset that simulates real-world scenarios for signature verification and detection. This dataset, expanded with various types of noise and forged signature data, is designed to test and improve the robustness of signature verification systems. The scope of the dataset reflects the challenges encountered in current research and applications, providing a valuable resource for developing new models and algorithms in this field.

Although access to a large number of Ottoman documents has become easier today, the Arabic-Persian-based Ottoman script remains a barrier for interested users in utilizing these documents. To address this challenge, there is a need for automatic transcription systems. While some deep learning-based commercial and academic models exist for Ottoman transcription, no studies have yet explored models based on transformer architectures. This paper introduces an Ottoman transcription system developed using TrOCR, a transformer-based model. Instead of the commonly used two-step approach in the literature, a model was designed to perform both optical character recognition and transcription into Turkish in one step. Additionally, the decoder responsible for language modeling was initialized with a BERT-based model trained on Turkish data, achieving results comparable to the original model. During testing, this model produced outputs more quickly due to improved tokenization performance.

A central challenge in representation learning is constructing latent embeddings that are both expressive and efficient. In practice, deep networks often produce redundant latent spaces where multiple coordinates encode overlapping information, reducing effective capacity and hindering generalization. Standard metrics such as accuracy or reconstruction loss provide only indirect evidence of such redundancy and cannot isolate it as a failure mode. We introduce a redundancy index, denoted rho(C), that directly quantifies inter-dimensional dependencies by analyzing coupling matrices derived from latent representations and comparing their off-diagonal statistics against a normal distribution via energy distance. The result is a compact, interpretable, and statistically grounded measure of representational quality. We validate rho(C) across discriminative and generative settings on MNIST variants, Fashion-MNIST, CIFAR-10, and CIFAR-100, spanning multiple architectures and hyperparameter optimization strategies. Empirically, low rho(C) reliably predicts high classification accuracy or low reconstruction error, while elevated redundancy is associated with performance collapse. Estimator reliability grows with latent dimension, yielding natural lower bounds for reliable analysis. We further show that Tree-structured Parzen Estimators (TPE) preferentially explore low-rho regions, suggesting that rho(C) can guide neural architecture search and serve as a redundancy-aware regularization target. By exposing redundancy as a universal bottleneck across models and tasks, rho(C) offers both a theoretical lens and a practical tool for evaluating and improving the efficiency of learned representations.

Learning a discriminative semantic space using unlabelled and noisy data remains unaddressed in a multi-label setting. We present a contrastive self-supervised learning method which is robust to data noise, grounded in the domain of variational methods. The method (VCL) utilizes variational contrastive learning with beta-divergence to learn robustly from unlabelled datasets, including uncurated and noisy datasets. We demonstrate the effectiveness of the proposed method through rigorous experiments with multi-label datasets in the face understanding domain, including one where the system is pretrained with web collected face images. Experiments include linear evaluation and fine-tuning scenarios, in addition to verification and face attribute learning tests, showing that the model learns effective embedding representations. In almost all tested scenarios, VCL surpasses the performance of state-of-the-art self-supervised methods.

AbstractAbstractive summarization is an approach to document summarization that is not limited to selecting sentences from the document but can generate new sentences as well. We address the two main challenges in abstractive summarization: how to evaluate the performance of a summarization model and what is a good training objective. We first introduce new evaluation measures based on the semantic similarity of the input and corresponding summary. The similarity scores are obtained by the fine-tuned BERTurk model using either the cross-encoder or a bi-encoder architecture. The fine-tuning is done on the Turkish Natural Language Inference and Semantic Textual Similarity benchmark datasets. We show that these measures have better correlations with human evaluations compared to Recall-Oriented Understudy for Gisting Evaluation (ROUGE) scores and BERTScore. We then introduce a deep reinforcement learning algorithm that uses the proposed semantic similarity measures as rewards, together with a mixed training objective, in order to generate more natural summaries in terms of human readability. We show that training with a mixed training objective function compared to only the maximum-likelihood objective improves similarity scores.

With the accelerated pace of digitization, a vast collection of Ottoman documents has become accessible to researchers and the general public. However, most users interested in these documents are unable to read them, as the text is Turkish written in the Arabic-Persian script. Manual transcription of such a massive amount of documents is also beyond the capacity of human experts. With the advancements in deep learning, we have been able to provide a solution to the long-standing problem of automatic transcription of printed Ottoman documents. We evaluated three decoding strategies including Word Beam Search that allows to use a recognition lexicon and n-gram statistics during the decoding phase. Furthermore, the effect of lexicon size and coverage and language modelling via character or word n-grams are also evaluated. Using a general purpose large lexicon of the Ottoman era (260K words and 86 6.59% character error rate and 28.46% word error rate on a test set of 6, 828 text lines.

Hate speech spread on social media can strongly affect people and societies, especially when it targets a specific group of people in terms of religion, culture, or a specific social situation, such as refugees. For this reason, the detection and elimination of hate speech in social networks have attracted the attention of natural language processing researchers in recent years. A competition is organized to benchmark progress in Turkish hate speech recognition, within SIU2023 Computational Social Sciences special session, with four different tasks. A total of 20 teams registered for the competition, while eight teams submitted results at the end. The details of the winning models and their results are explained in this paper.