News

Dr. Algin and his team’s new study about WEB-assisted aneurysm embolization has been published.

The 2020 and 2021 IEEE Turkey Science Awards were presented to their owners at the ceremony held at Bilkent University on December 18, 2021. UMRAM faculty member Emine Ülkü Sarıtaş was deemed worthy of the 2020 IEEE Turkey Research Incentive Award.

Parlar foundation gives Honorary, Science, Service, Research and Technology Incentive Awards every year in order to evaluate the research studies and services carried out in science and industry and to encourage the growing generations. Hulusi Kafalıgönül, faculty member at UMRAM, has received 2021 Parlar Research Incentive Award. Dr.Kafalıgönül’s research group at Bilkent University carries out studies using several different techniques on sensory processing, multi-sensory integration, and perception dynamics from a systemic neuroscience perspective. These studies are supported by TUBITAK and other international (EU Marie Curie Actions, British Academy) research institutions.

Emin Çelik has presented his PhD thesis titled ”Spatially Informed Voxelwise Modeling and Dynamic Scene Category Representation in the Human Brain” on December 15, 2021.   ABSTRACT Humans have an impressive ability to rapidly process global information in natural scenes to infer their category. Yet, it remains unclear whether and how scene categories observed dynamically in the natural world are represented in cerebral cortex beyond few canonical scene-selective areas. To address this question, here we examined the representation of dynamic visual scenes by recording whole-brain blood oxygenation level-dependent (BOLD) responses while subjects viewed natural movies. We fıt voxelwise encoding models to estimate tuning for scene categories that reflect statistical ensembles of objects and actions in the natural world. Voxelwise modeling (VM) is a powerful framework to predict single voxel responses evoked by a rich set of stimulus features present in complex natural stimuli. However, because VM disregards correlations across neighboring voxels, its sensitivity in detecting functional selectivity can be diminished in the presence of high levels of measurement noise. Here, we introduce spatially-informed voxelwise modeling (SPIN-VM) to take advantage of response correlations in spatial neighborhoods of voxels. To optimally utilize shared information, SPIN-VM performs regularization across spatial neighborhoods in addition to model features, while still generating single-voxel response predictions. Compared to VM, SPIN-VM yields higher prediction accuracies and better capture locally congruent information representations across cortex. We fınd that this scene-category model explains a signifıcant portion of the response variance broadly across cerebral cortex. Cluster analysis of scene-category tuning profıles across cortex reveals nine spatially-segregated networks of brain regions consistently across subjects. These networks show heterogeneous tuning for a diverse set of dynamic scene categories related to navigation, human activity, social interaction, civilization, natural environment, non-human animals, motion-energy, and texture, suggesting that the organization of scene category representation is quite complex.   Keywords: fMRI, Voxelwise modeling, Computational neuroscience, Spatial regularization, Coherent representation, Dynamic scene category representation, Cluster analysis. Supervisor: Assoc. Prof. Dr. Tolga Çukur

Assoc. Prof. Tolga Çukur of the Department of Electrical and Electronics Engineering and UMRAM has received an award from TÜSEB (the Health Institutes of Turkey). Named after Nobel laureate Aziz Sancar, the TÜSEB awards honor outstanding work by the country’s researchers in fields related to the health sciences. Dr. Çukur is the recipient of a TÜSEB Aziz Sancar Incentive Award. These awards are given to a maximum of five recipients each year: scientists under the age of 40 who have proven their potential to contribute to science at an international level. Dr. Çukur received this award for his contributions in the field of biomedical imaging. His research topics include machine learning techniques for data acquisition and processing on next-generation computational imaging platforms, and radiological and neuroscientific applications of biomedical imaging.

To tackle the problem of lexical semantic change, KocLab has developed the first Gaussian word embedding (w2g)-based approach, which has been described in the paper “Semantic Change Detection with Gaussian Word Embeddings” published at IEEE Transactions on Audio, Speech and Language Processing. The proposed method reached high rankings (including the first rank at one of the two sub-tasks) at the SemEval 2020 Task 1, which is the standardized evaluation framework of the field. Languages evolve with time since cultural and linguistic effects alter meanings of words in a semantic space and diachronic study of semantic change explores those changes in word meanings.  Computationally detecting and analyzing lexical semantic change are of fundamental importance since, let alone the theoretical importance for computational semantics and linguistics, identifying semantically altered words improves the performances of a wide range of high-level NLP applications. KocLab makes a significant contribution to the field of lexical semantic change detection by proposing this theoretically motivated approach utilizing w2g embeddings that results in high correlation with human judgements. In addition to modelling relative positions of words within semantic vector spaces, the variance components of w2g embeddings model semantic coverages of words. This aspect of w2g-based models provides crucial information since semantic change occurs not only as a drift of meaning within semantic space but also as narrowing or broadening of semantic coverage. With the successful introduction of w2g-based models to the diachronic NLP studies, we expect to see future efforts exploring ways to exploit word variances in LSC detection and categorization. The paper can be accessed [here].   Abstract: Diachronic study of the evolution of languages is of importance in natural language processing (NLP). Recent years have witnessed a surge of computational approaches for the detection and characterization of lexical semantic change (LSC) by using advancing word representation techniques and the availability of diachronic corpora. We propose a Gaussian word embedding (w2g) based methodology and present a comprehensive study for the LSC detection. W2g is a probabilistic distribution-based word embedding model and represents words as Gaussian mixture models using covariance information along with the existing mean (word vector). We also extensively study several aspects of w2g-based LSC detection under the SemEval-2020 Task 1 evaluation framework as well as using Google N-gram corpus. In the Sub-task 1 (LSC binary classification) of the SemEval-2020 Task 1, we report the highest overall ranking as well as the highest ranks for the two (German and Swedish) of the four languages (English, Swedish, German and Latin). We also report the highest Spearman correlation in the Sub-task 2 (LSC ranking) for Swedish. Our overall rankings in the LSC classification and ranking sub-tasks are 1st and 7th, respectively. Qualitative analysis has also been presented.

The work of Dr. Koç and his collaborators on sentiment analysis has been published in IEEE Transactions on Neural Networks and Learning Systems. The paper entitled “Multi-Label Sentiment Analysis on 100 Languages with Dynamic Weighting for Label Imbalance” proposes a novel method to overcome the inherent label imbalance problem in multi-label classification by balancing losses from different classes. The proposed model can classify sentiment in 100 different languages and achieves first place in 7 of 9 metrics in English, Arabic and Spanish languages in the SemEval competition using only a single model.   The paper can be accessed [here].   Abstract: We investigate cross-lingual sentiment analysis, which has attracted significant attention due to its applications in various areas including market research, politics, and social sciences. In particular, we introduce a sentiment analysis framework in multi-label setting as it obeys Plutchik’s wheel of emotions. We introduce a novel dynamic weighting method that balances the contribution from each class during training, unlike previous static weighting methods that assign non-changing weights based on their class frequency. Moreover, we adapt the focal loss that favors harder instances from single-label object recognition literature to our multi-label setting. Furthermore, we derive a method to choose optimal class-specific thresholds that maximize the macro-f1 score in linear time complexity. Through an extensive set of experiments, we show that our method obtains the state-of-the-art performance in seven of nine metrics in three different languages using a single model compared with the common baselines and the best performing methods in the SemEval competition. We publicly share our code for our model, which can perform sentiment analysis in 100 languages, to facilitate further research.

The paper entitled “Operator Theory-based Computation of Linear Canonical Transforms” published in Signal Processing by Dr. Koç and his collaborators applied the operator theory to the computation of linear canonical transforms. By exploiting hyperdifferential operators and through a systematic investigation of possible parameters and design choices, the proposed algorithm achieves a discrete linear canonical transform (DLCT) definition that is both theoretically satisfying and highly accurate. The advantage and elegance of this approach lie in the fact that it reduces the problem of defining sophisticated discrete transforms to much simpler problem of defining discrete coordinate multiplication and differentiation operations. The proposed approach may lead to further possible research directions in the theory of discrete transforms. The paper can be accessed [here].   Abstract: Linear canonical transforms (LCTs) are extensively used in many areas of science and engineering with many applications, which requires a satisfactory discrete implementation. Recently, hyperdifferential operators have been proposed as a novel way of defining the discrete LCT (DLCT). Here we first focus on improving the accuracy of this approach by considering alternative discrete coordinate multiplication and differentiation operations. We also consider canonical decompositions of LCTs and compare them with the originally proposed Iwasawa decomposition. We show that accuracy of the approximation of the continuous LCT with the DLCT can be drastically improved. The advantage and elegance of this approach lie in the fact that it reduces the problem of defining sophisticated discrete transforms to merely defining discrete coordinate multiplication and differentiation operations, by reducing the transforms to these operations. As a result of systematic investigation of possible parameters and design choices, we achieve a DLCT that is both theoretically satisfying and highly accurate.  

A new study of Dr. ALGIN and his team has been published.

Prof. Dr. Oktay ALGIN and his colleagues recently published the results of a study entitled ‘Cerebrospinal fluid velocity changes in idiopathic scoliosis: A preliminary study using 3-tesla PC-MRI and 3D-SPACE-VFAM data’ in the scientific journal “Child’s Nervous System.”

The findings of this study indicate that people with idiopathic scoliosis have hyperdynamic or altered spinal cerebrospinal fluid flow. These findings can be used to initiate or open the way for additional research, such as predicting and/or preventing spinal curve progression and examining the impact of various conservative and surgical treatment procedures using cerebrospinal fluid flow analysis. As a result, the findings of the study have clinical implications.

3-tesla magnetic resonance imaging (MRI) will continue to be used in clinical trials to assess spinal cerebrospinal fluid.

The related article can be accessed at the web address below;

https://link.springer.com/journal/381/online-first

As an outcome of TÜBİTAK 1001 Project named “Natural Language Processing and Machine Learning Techniques for Applications of Artificial Intelligence in Legal Texts”, KocLab is pleased to announce that a journal paper has been published in Information Processing & Management. The paper “Natural Language Processing in Law: Prediction of outcomes in the Higher Courts of Turkey” is a pioneering work that addresses the Turkish Legal System in a comprehensive manner and develops Natural Language Processing (NLP) techniques for the legal system of the Republic of Turkey. The problem of predicting the decisions of the Turkish higher courts (Constitutional Court, Civil Court of Appeal, Criminal Court of Appeal, Administrative Court of Appeal and Court of Appeal on Taxation) are tackled with the use of various Machine Learning (ML) algorithms. The decisions of the courts are predicted using only fact descriptions. Specific to the Turkish legal system, codified in Turkish, several algorithms utilizing Decision Trees (DTs), Random Forests (RF), Support Vector Machines (SVM) and 3 state-of-the-art Deep Learning (DL) algorithms (Gated Recurrent Units (GRUs), Long Short-Term Memory networks (LSTMs) and bidirectional LSTMs (BiLSTMs) are developed. An attention mechanism specific to legal texts has also been integrated into the DL algorithms. This study is broader than the similar works in other legal systems and languages in literature in terms of the number of investigated courts and the number of utilized algorithms. The study also discusses some of the practical applications, and the legal and ethical implications of the use of such machine-based predictive systems. Hence, it not only pioneers the NLP and AI for our country’s legal system but also provides a reference point and baseline for further studies in this field. The paper can be accessed here. Abstract: Natural language processing (NLP) based approaches have recently received attention for legal systems of several countries. It is of interest to study the wide variety of legal systems that have so far not received any attention. In particular, for the legal system of the Republic of Turkey, codified in Turkish, no works have been published. We first review the state-of-the-art of NLP in law, and then study the problem of predicting verdicts for several different courts, using several different algorithms. This study is much broader than earlier studies in the number of different courts and the variety of algorithms it includes. Therefore it provides a reference point and baseline for further studies in this area. We further hope the scope and systematic nature of this study can set a framework that can be applied to the study of other legal systems. We present novel results on predicting the rulings of the Turkish Constitutional Court and Courts of Appeal, using only fact descriptions, and without seeing the actual rulings. The methods that are utilized are based on Decision Trees (DTs), Random Forests (RFs), Support Vector Machines (SVMs) and state-of-the-art deep learning (DL) methods; specifically Gated Recurrent Units (GRUs), Long Short-Term Memory networks (LSTMs) and bidirectional LSTMs (BiLSTMs), with the integration of an attention mechanism for each model. The prediction results for all algorithms are given in a comparative and detailed manner. We demonstrate that outcomes of the courts of Turkish legal system can be predicted with high accuracy, especially with deep learning based methods. The presented results exhibit similar performance to earlier work in the literature for other languages and legal systems.

Dr. Manouchehr Takrimi’s work on computational electromagnetics has recently been published in IEEE Transactions on Antennas and Propagation. Incomplete tree structures are used in conjunction with the mixed-form multilevel fast multipole algorithm (MLFMA) for the solution of potential integral equations (PIEs) leading to an algorithm that can perform an accurate and efficient analysis of electromagnetic scattering/radiation for multiscale open and closed surfaces. The presented algorithm is also population based and deploys nonuniform clustering for multiscale geometries that are often modeled by using highly variable discretizations. However, it is significantly modified from its original version based on surface integral equations (SIEs) published in 2016, and its modified version for volume integral equations (VIEs) published in 2017. In the recent work, by introducing pruned boxes, and reorganizing box-to-box as well as near- and far-field interactions, the necessary means are provided in order to be combined with PIEs. Numerical results obtained from the scattering/radiation analysis of canonical and real-life multiscale geometries clearly demonstrate the superiority of the algorithm. It is accurate, where the error is controllable, and is very efficient in terms of run-time and especially memory requirements, where more than an order of magnitude saving in memory can be achieved.   The paper can be accessed here.   Abstract:  Recently introduced incomplete tree structures for the magnetic-field integral equation are modified and used in conjunction with the mixed-form multilevel fast multipole algorithm (MLFMA) to employ a novel broadband incomplete-leaf (IL) MLFMA (IL-MLFMA) to the solution of potential integral equations (PIEs) for scattering/radiation from multiscale open and closed surfaces. This population-based algorithm deploys a nonuniform clustering that enables to use deep levels safely and, when necessary, without compromising the accuracy resulting in an improved efficiency and a significant reduction for the memory requirements (order of magnitudes) while the error is controllable. The superiority of the algorithm is demonstrated in several canonical and real-life multiscale geometries.

Dr. Aykut Koç’s work on graph signal processing (GSP) has recently been published in IEEE Transactions on Signal Processing. In this contribution to the graph signal processing (GSP), Dr. Koç and his collaborators formulate the problem of optimal filtering in fractional Fourier domains on graphs and provide closed form solution of it for GSP. We generalize the optimal Wiener filtering problem on graphs to fractional Fourier domains. This is the generalization of the corresponding classical problem in ordinary Fourier domains, and has the potential to offer improved performance in the sense of lower errors, thanks to the additional degree of freedom in choosing the fractional Fourier transform order. The optimal fractional Fourier domain filtering problem has been theoretically generalized to GSP framework and several numerical applications in which filtering in fractional domains yields smaller error values than filtering in the vertex-frequency domain have been demonstrated. Neither on the graph structure nor on the graph signals, did we make any restrictive assumptions. Therefore, the proposed fractional graph Fourier transform filtering framework can be applied under a wide variety of arbitrary conditions. The paper can be accessed [here].   Abstract: Graph signal processing has recently received considerable attention. Several concepts, tools, and applications in signal processing such as filtering, transforming, and sampling have been extended to graph signal processing. One such extension is the optimal filtering problem. The minimum mean-squared error estimate of an original graph signal can be obtained from its distorted and noisy version. However, the best separation of signal and noise, and thus the least error, is not always achieved in the ordinary Fourier domain, but rather a fractional Fourier domain. In this work, the optimal filtering problem for graph signals is extended to fractional Fourier domains, and theoretical analysis and solution of the proposed problem are provided along with computational cost considerations. Numerical results are presented to illustrate the benefits of filtering in fractional Fourier domains.