In this thesis study, we focus on the construction of an effective network architecture, for which we propose an architecture generation framework and show how it can be used to create an effective terrain classification model. Additionally, we also observe that land cover training data sets on HSI and LiDAR tend to come short in providing training examples with shadow effects. To address this limitation, we additionally propose a generative adversarial network(GAN) driven statistical data augmentation technique that generates synthetic training examples and show its effectiveness in our experimental results.

Date: 26.01.2023 / 14:30 Place: A-108

This thesis demonstrates research towards innovating a novel framework that (i) could be trained by supervised learning of discourse parsing from annotated and parsed relations, (ii) makes it easy to build end-to-end shallow discourse parsing system, (iii) forms a discourse relation detector benchmark for low resourced languages. Our modern neural network approach, integrated with contextualized text embedding, produced by pretrained language models, accomplished two sub-tasks of shallow discourse parsing, namely, identification of discourse relation realization types and the sense classification of explicit and implicit relations. The effect of multilingual data aggregation on the classification of discourse relation type through Cross-lingual Transfer Learning experiments is researched too.

Date: 25.01.2023 / 10:00 Place: A-108

Building on an extensive review of the psycholinguistics literature and Turkish Derivational Morphology (DM), we propose a novel structure for representing DM in three hierarchical layers: segmentation, lexical selection and derivation. This proposal involves laying a conventionalized structure over the traditional morphological structure of DM. We develop a computational model of morphology processing based on this structure using Bayesian Belief Networks (BBN). We present an algorithmic implementation for this model that learns and accurately represents new lexical items, recognizes affixes and tracks the salience of each item probabilistically. We carry out trials on this model with realistic observation lists and observe that model predictions are in line with the findings in studies in psycholinguistics.

Date: 25.01.2023 / 12:00 Place: A-108

Musical conducting is the art of making real time manipulations on a musical piece as if the instruments are the musicians themselves. Orchestra responses affect the conductor’s actions while a musical piece is being played. Therefore there is a significant information flow on the stage. In this research, we focused on the causality relations between the conductor, orchestra and the musical scores. We collected data from an actual orchestral practice and processed the collected data for causality measurements using the Transfer Entropy method. We utilized the method also for unsupervised detection of conducting commands.

Date: 20.01.2023 / 10:00 Place: A-108

In Türkiye, Social Security Institution is the primary healthcare insurer. Turkish citizens are registered under General Medicare Insurance coverage. In 2003, Ministry of Health (MoH) has initiated the “Health Transformation Program”, and implemented the interoperable health data exchange standards. The MoH is focusing on collecting medical data in a coded, structured, and electronic format, generated at all healthcare providers. Contrarily, genetic test results are exchanged in narrative, unstructured form among governmental and private health care providers. In this dissertation, we lay out the bottlenecks and put forward a conceptual model for meaningful genomic data exchange for Turkish Electronic Health Records.

Date: 18.01.2023 / 15:00 Place: B-116

The fusion of information coming from different sensors (e.g., optics, audio, accelerometer) to recognize egocentric activities is still an active research area. Although the increase in sensor diversity brings out the need for adaptive fusion, there is a limited number of studies. In this work, we proposed two novel multi-modal decision fusion frameworks for egocentric activity recognition. The first framework combines hand-crafted features using Multi-Kernel Learning. The other framework utilizes deep features using a two-stage decision fusion mechanism. Additionally, a new egocentric activity dataset, named Egocentric Outdoor Activity Dataset (EOAD), was populated, containing 30 egocentric activities and 1392 video clips.

Date: 18.01.2023 / 14:00 Place: A108

A new method has been developed to study allosteric protein regulation, which is important for understanding how proteins function. The method represents proteins as networks and identifies allosteric pathways, or sites where molecules bind and regulate protein activity, in order to predict the presence and location of allosteric regions. This could potentially aid in the development of targeted therapies.

Date: 18.01.2023 / 10:00 Place: A212

The software industry is playing a significant role in development of economies all over the world. It is mainly made up of small and medium software enterprises (SMEs). These companies aim to benefit from Software Process Improvements (SPI) to increase product quality and productivity. As SPI require organizational change and adaptation to new work practices; organizations have to handle with several challenges emerged from this change. The objective of this study is to identify the factors that influence the success and the acceptance of SPI models used in SMEs. Moreover, it is aimed to analyze the attitude of SME employees towards the SPI models in Turkey.

Date: 25.01.2023 / 14:00 Place: B-116

The statistical analysis of the microbiome data obtained by the Next Generation Sequencing technology assumes that the data lies on the real space. Since 2017, it has been argued that microbiome data analyzes should be performed on simplex, that is a subspace of real space. Because, microbiome data is "compositional". In this thesis, the characteristics of microbiome data and the statistical difficulties of microbiome analysis are examined. Then, a new grouping method (Principal Microbial Groups) was introduced, using the compositional nature of the data, which is an alternative to the phylogenetic grouping of microbiome data.

Date: 18.11.2022 / 11:30 Place: A212

In deep learning, small input shifts or translations can cause dramatic changes in the output. This is because of the fact that commonly used down-sampling techniques such as max pooling, strided convolution, and average pooling ignore the sampling theorem. We have demonstrated that the cycle-spinning (CS) signal processing technique can be used before down-sampling in deep learning to increase accuracy without introducing any extra learnable parameters. The proposed method can be applied to different algorithms such as GAN, classification, and object detection.

Date: 29.11.2022 / 13:00 Place: METU Informatics Institute