10
Arşivlenen Tez
1
DOI Atanmış
10%
DOI Oranı
Anabilim Dalı
indexliAçık ErişimTRAğ trafiği kullanarak saldırı tespit edebilen derin öğrenme modellerinin performans değerlendirmesi
Dijital altyapıların hızla büyümesi ve siber saldırıların giderek karmaşıklaşması, ağ tabanlı saldırı tespit sistemlerini modern siber güvenliğin temel bileşenlerinden biri hâline getirmiştir. İmza tabanlı yöntemler yeni tehditlerin tespitinde yetersiz kalabildiğinden, makine öğrenmesi ve derin öğrenme yöntemleri önemli alternatifler sunmaktadır. Bu çalışmada, tek boyutlu evrişimli sinir ağı (CNN) ile hibrit CNN-BiLSTM mimarisinin ikili ve çok sınıflı ağ saldırısı sınıflandırma performansları incelenmiştir. Modeller, CIC-IDS2017 ve UNSW-NB15 veri kümeleri üzerinde tabakalı beş katlı çapraz doğrulama, odak kaybı, QuantileTransformer ile ölçeklendirme ve SMOTE ile aşırı örnekleme içeren ortak bir deneysel düzen kullanılarak değerlendirilmiştir. Ayrıca rastgele orman ve XGBoost, UNSW-NB15 üzerinde klasik karşılaştırma modelleri olarak test edilmiştir. Bulgular, CNN modelinin CIC-IDS2017 ikili sınıflandırma görevinde %99,75 doğruluk ve 0,9999 ROC-AUC değerine ulaştığını göstermiştir. CNN-BiLSTM, UNSW-NB15 ikili sınıflandırma görevinde yanlış negatifleri yaklaşık %36 azaltmış; ancak çıkarım gecikmesi CNN’e göre 4-6 kat artmıştır. Çok sınıflı sınıflandırmada CNN, CIC-IDS2017 üzerinde 0,701 makro F1 skoruyla CNN-BiLSTM’den (0,671) daha yüksek performans göstermiştir. XGBoost ise derin öğrenme modelleriyle benzer F1 skorlarına daha kısa sürede ulaşmıştır. Azınlık saldırı sınıflarının tespiti tüm yapılandırmalarda temel bir güçlük olmayı sürdürmektedir. Sonuçlar, gerçek uygulama kısıtları altında derin öğrenme ve klasik makine öğrenmesi modelleri arasında seçim yapacak araştırmacı ve uygulayıcılar için karşılaştırmalı çıkarımlar sunmaktadır.
Verimli ve etkili öğrenme için rezervuar hesaplama
Resource-efficient modeling of nonlinear dynamical systems is vital for embedded sensing, edge analytics, and neuromorphic hardware. Reservoir Computing (RC) reduces training overhead in recurrent neural networks and transformers by fixing internal weights and training only a linear read-out. However, classical Echo State Networks (sRC) scale capacity by increasing reservoir size, which inflates memory and energy costs. This thesis addresses that bottleneck by improving the quality, rather than the quantity, of reservoir units. We propose the Higher-Order Augmented Reservoir Computer (haRC), which augments reservoir with a subset of multiplicative monomials of co-temporal activations. This polynomial feature space enhances nonlinear expressivity while keeping the reservoir compact. On chaotic benchmarks such as Lorenz and Mackey–Glass, haRC achieves the same or better accuracy with fewer parameters than baseline sRC models.We propose two improved models: Selective haRC, which ranks reservoir units by variance and covariance to discard redundant interactions, and Sparse haRC, which integrates controlled sparsity masks to reduce memory and compute load. Comprehensive experiments evaluate memorization and forecasting tasks under equal total and tunable parameter settings across different sparsity levels. haRC's consistently outperform sRC across all sparsity regimes, with the advantage growing in high sparsity. Noise robustness tests inject Gaussian noise into the reservoir. Selective haRC maintains or exceeds the performance of sRC in terms of root-mean-squared error and variance with 12% of the reservoir units sRC has. Demonstrating up to 120 times memory efficiency, higher-order augmentation emerges as a promising approach for ultra-lightweight sequence learners suited for microcontrollers, FPGA overlays, and in-memory compute systems.
SQTL ve HI-C'nin ortak analizi, birden fazla dokuda Sqtller ile hedef genler arasında mekansal yakınlık ortaya çıkarıyor
Gene expression and regulation with or without alternative splicing are crucial for tissues and cells to properly function. They have been studied from three almost independent perspectives at the genome level: 1- Recognition of splicing quantitative trait loci~(sQTLs), 2- Expression quantitative trait loci~(eQTLs) recognition, and 3- Recognition of longer-range physical chromatin interactions between genome segments which model $3$D dynamics of cells and tissues. Even though the associations between eQTLs and longer range chromatin interactions have been previously studied, similar relationship between sQTLs and chromatin interactions has not been previously analyzed. In this case, it is crucial to analyze whether sQTLs control the alternative splicing of their target genes mRNA via physically-interacting genome segments. Even though chromatin interactions are part of the principal processes governing eQTLs functioning, similar analysis is missing from sQTLs perspective. We have jointly analyzed high-throughput chromatin conformation capture~(Hi-C) and sQTL datasets over 8 different human cancer tissues. We have discovered the existence of positive association between the number of genes having sQTLs and chromatin interaction frequency. Such positive association still exists when we also control for eQTLs. Additionally, sQTLs and their target genes generally exist inside identical topologically associating domain~(TAD). Those findings are observed over the whole set of analyzed cancer types and over different functional subsets of sQTL dataset such as survival-related sQTLs. Furthermore, tissue-specific sQTLs are statistically enriched in tissue-specific frequently interacting regions~(FIREs) in 6 out of 8 human cancer tissues~(Chronic Myeloid Leukemia, Colon Adenocarcinoma, Acute Myeloid Leukemia, Lung Adenocarcinoma, Prostate Cancer, Sarcoma). Our sQTL and Hi-C datasets have shown the existence of closer spatial distance between sQTLs and their target genes with possible alternative splicing across a number of different cancer types in human. Such closer spatial distance also exists independent of whether we integrate eQTLs into the analysis. We found that sQTLs regulate the alternative splicing through chromatin interactions.
Kombinatoryal optimizasyon için grafik sinir ağları tabanlı birincil sezgisel yöntem
By examining the patterns of solutions obtained for varying instances, one can gain insights into the structure and behavior of combinatorial optimization (CO) problems and develop efficient algorithms for solving them. Machine learning techniques, especially Graph Neural Networks (GNNs), have shown promise in parametrizing and automating this laborious design process. The inductive bias of GNNs allows for learning solutions to mixed-integer programming (MIP) formulations of constrained CO problems with a relational representation of decision variables and constraints. The trained GNNs can be leveraged with primal heuristics to construct high-quality feasible solutions to CO problems quickly. However, current GNN-based end-to-end learning approaches have limitations for scalable training and generalization on larger-scale instances; therefore, they have been mostly evaluated over small-scale instances. Addressing this issue, our study builds on end-to-end learning of optimal solutions to the downscaled instances of given large-scale CO problems. We introduce several improvements on a recent GNN model for CO to generalize on instances of a larger scale than those used in the training. We also propose a two-stage primal heuristic strategy based on uncertainty-quantification to automatically configure how solution search relies on the predicted decision values. Our models can generalize on 16x upscaled instances of commonly benchmarked five CO problems. Unlike the regressive performance of existing GNN-based CO approaches as the scale of problems increases, the CO pipelines using our models offer an incremental performance improvement relative to a state-of-the-art MIP solver CPLEX. The proposed uncertainty-based primal heuristics provide 6-75% better optimality gap values and 45-99% better primal gap values for the 16x upscaled instances and brings immense speedup to obtain high-quality solutions. All these gains are achieved in a computationally efficient modeling approach without sacrificing solution quality.
A comprehensive human-agent negotiation framework: Preferences, emotion & interaction
In today's increasingly interconnected world, human-agent negotiation plays a pivotal role in reaching socially beneficial agreements when stakeholders need to make joint decisions. Developing intelligent agents capable of understanding not only human negotiators' preferences but also attitudes is a significant prerequisite for effective human-agent interactions. Awareness of a human's emotional state and ability to express an agent's mood to influence the human negotiator might significantly affect the negotiation outcome. This thesis presents a comprehensive framework that revolutionizes the field of human-agent negotiation, integrating two critical elements: Emotionally aware negotiation strategy and Conflict-Based Opponent Modeling (CBOM). By combining these novel approaches, the framework enhances negotiation outcomes and fosters cooperation between agents and human negotiators, ultimately leading to mutually advantageous agreements. The thesis establishes the research context and motivation, underscoring the escalating importance of human-agent negotiation in a world where collaborative decision-making is essential for addressing complex challenges. It highlights the need for advanced agents to accurately interpret human preferences and behaviors, enabling admissible settlements that serve joint interests. Shedding light on the limitations of conventional approaches that heavily rely on opponent offers and remaining time. Additionally, it explores the critical role of emotional awareness and opponent modeling strategies in human-agent negotiation. The synthesis of existing research lays the groundwork for developing the proposed comprehensive framework. Emotional awareness takes center stage in the proposed negotiation strategy. Solver Agent: Emotional Extension of the Hybrid Agent bidding strategy is introduced. The Solver Agent considers the opponent's emotional state during negotiation, leading to higher social welfare scores and faster agreement times. The experimental study emphasized the profound impact of emotional awareness on negotiation outcomes, particularly in human-agent settings. CBOM efficiently extracts maximum information from limited interaction rounds in human-agent negotiation settings, surpassing traditional approaches in prediction performance. Experimental analyses confirmed the superiority of CBOM in human-agent and automated negotiation scenarios, even when the exploration of the outcome space is limited. The experimental findings establish CBOM as a powerful tool for modeling human behavior and preferences in negotiation. In conclusion, the comprehensive human-agent negotiation framework presented in this thesis represents a significant advancement in the field. By seamlessly combining Conflict-Based Opponent Modeling and Emotional Awareness, the framework empowers intelligent agents to discern human preferences and behaviors more accurately, facilitating cooperative interactions and achieving mutually beneficial agreements. The framework's effectiveness in human-agent and automated negotiation settings highlights its potential for designing negotiation agents that interact adeptly with human negotiators, fostering understanding and optimizing negotiation outcomes. The future of human-agent negotiation lies in forging a new era of cooperation, where intelligent agents serve as capable partners, promoting social welfare and driving positive change through admissible settlements that incorporate joint interests. This thesis contributes valuable insights towards realizing this vision, marking a significant step forward in the field of human-agent interaction.
Illumination-guided inverse rendering benchmark: Learning real objects with few cameras
The field of 3D computer vision and graphics has seen a rapid expansion of late, making the development of realistic virtual environments and digital representations of real-world objects possible. Fundamental to this development are 3D reconstruction techniques that enable the transition of physical objects' form, color, and surface particulars to the virtual domain. Current approaches mainly rely on neural scene representations, which, despite being effective, face challenges such as the need for a large quantity of captured images and the complexity associated with converting these representations into explicit geometric forms. An alternative strategy that has gained traction is the deployment of methods such as physically-based differentiable rendering (PBDR) and inverse rendering. These approaches require fewer viewpoints, yield explicit format results, and ensure a smoother transition to other representation methods. However, in order to effectively assess the performance of the available methods in 3D reconstruction, it is imperative to utilize standard benchmark scenes for comparison. Although there are standard objects and scenes available in existing research, there is a noticeable deficiency of real-world benchmark data that simultaneously captures camera, lighting, and scene parameters — all of which are essential for high-quality 3D reconstructions using methods based on PBDR and inverse rendering. In this study, we present a method for capturing real-world scenes as virtual environments, incorporating lighting parameters along with camera and scene parameters to enhance the veracity of virtual representations. In addition, we provide a set of ten real-world scenes, each with its corresponding virtual counterparts, purposely designed as benchmarks. These benchmarks cover a basic assortment of geometric structures, such as convex, concave, flat, and composite surfaces. Furthermore, we showcase the 3D reconstruction results of cutting-edge 3D reconstruction techniques using PBDR in real-world scenes, using both established methodologies and our proposed one.
Derı̇n transformatör tabanlı varlık fı̇yatı ve yön tahmı̇nı̇
The field of algorithmic trading, driven by deep learning methodologies, has garnered substantial attention in recent times. Within this domain, transformers, convolutional neural networks, and patch embedding-based techniques have emerged as popular choices within the computer vision community. Here, inspired by the latest cutting-edge computer vision methodologies and the existing work showing the capability of image-like conversion for time-series datasets, we apply more advanced transformer-based and patch-based approaches for predicting asset prices and directional price movements. The employed transformer models include Vision Transformer~(ViT), Data Efficient Image Transformers~(DeiT), and Swin. We use ConvMixer for a patch embedding-based convolutional neural network architecture without a transformer. Our tested transformer-based and patch-based methodologies aim to predict asset prices and directional movements using historical price data by leveraging the inherent image-like properties within the historical time-series dataset. Before the implementation of attention-based architectures, the historical time series price dataset is transformed into two-dimensional images. This transformation is facilitated through the incorporation of various common technical financial indicators, each contributing to the data for a fixed number of consecutive days. Consequently, a diverse set of two-dimensional images is constructed, reflecting various dimensions of the dataset. Subsequently, the original images depicting market valleys and peaks are annotated with labels such as Hold, Buy, or Sell. According to the experiments, trained attention-based models consistently outperform the baseline convolutional architectures, particularly when applied to a subset of frequently traded Exchange-Traded Funds~(ETFs). This better performance of attention-based architectures, especially ViT, is evident in terms of both accuracy and other financial evaluation metrics, particularly during extended testing and holding periods. These findings underscore the potential of transformer-based approaches to enhance predictive capabilities in asset price and directional forecasting. Our code and processed datasets are available at~https://github.com/seferlab/price_transformer
İkili müzakerelerde zaman serisi tahmin modelleri
This thesis explores the dynamics of agent-based negotiations, with a focus on understanding opponent's behavior and predicting their offering patterns to make strategic decisions. Guessing the utility of the opponent's upcoming offers valuable insights for the agent's subsequent moves. The research aims to predict the opponent's future offers by employing diverse learning algorithms in various situations to measure their effectiveness in comprehending negotiation behavior. The prediction study comprises two parts; one investigating the impact of these models in one-to-one negotiations, specifically tailored for the agent's own experiences, while the other examines the performance of predictive models in a tournament setting for all agents. A learning process with three distinct targets have been established to assess the prediction models: (i) estimating the agent's utility of the opponent's next offer by considering only its offer history, (ii) estimating the agent's utility considering opponent-related variables, and (iii) estimating the opponent's utility using opponent-related variables. According to the experimented results, the best learning approach is incorporated into an agent design to observe the impacts of having predictions of future utility values on the agent's negotiation success. The thesis evaluates these models in diverse negotiation scenarios and highlights promising outcomes for the proposed methods. It also introduces a novel negotiation strategy called `Negoformer', which incorporates predictions into the offering strategy and investigates their impact on the outcome of negotiations. The experiments showcased the success of Negoformer compared to other agents in various negotiation success metrics, such as individual utility value and social welfare score.
Gösterimden öğrenme için etkili sinir ağları
Learning from demonstration (LfD) is a powerful technique for teaching robots new behaviors by mimicking human demonstrations. One widely used approach in LfD is Behavior Cloning (BC) with human-in-the-loop control. In this method, data collected from humans demonstrations is utilized to create a non-linear controller by learning a mapping from states to desired actions. In this study, we propose a novel BC system that its structure mirrors the structure of an error-based feedback controller. This design choice is based on the premise that embedding such a structure within the learning model can endow our system with a prior bias, resulting in an inherent advantage over controller-agnostic BC systems. This thesis details the components of this novel BC model and demonstrates its application on a two degrees-of-freedom robotic system across various tasks as a proof of concept. To assess the system's effectiveness, we conducted systematic experiments comparing it to a controller-agnostic BC system. The results reveal that our proposed model significantly outperforms the baseline, suggesting it is a promising candidate for LfD tasks where the demonstration data can be assumed to be generated by a feedback controller, especially in resource-scarce conditions.
Akademik kalite ölçümü için yapay zeka tabanli bir yaklaşim
In the context of academic recruitment at universities and research institutions, establish- ing consistent and effective evaluation criteria remains a complex challenge. Identifying robust metrics that align with globally recognized standards of academic quality is essen- tial to ensure merit-based evaluation of researchers and maintain institutional credibility. In this thesis, we address this challenge through an AI-based solution aimed at developing a data-driven approach to quantify academic quality. As a benchmark of academic excellence, we use Nobel laureates' profiles in Physics, Chemistry, Physiology or Medicine, and Economics as a reference cohort. Comparison group includes researchers from the same fields affiliated with universities with an average ranking based on the Times Higher Education World University Rankings. By defining bibliometric features of the academic profiles of both Nobel laureates and the comparison group, we aim to develop machine learning models to quantify academic quality and identify the key features that define academic excellence. The ultimate goal is to support the decision-making process of universities and research institutions in academic recruitment, creating a fair and objective evaluation criteria.