Research

Interpretable Feature Selection for BCI

The goal is to develop a filter-based feature selection method by considering the neighborhood relationships and physiological attributes (such as regions of interest related to mental tasks) present in EEG data. The primary aim is to reduce temporal complexity, while the secondary objective is to select generalized and physiologically interpretable feature subsets by avoiding reliance on machine learning-based selection methods. In the field of artificial learning, interpretability studies generally focus on three main approaches: a) application-level, b) human-level, and c) functional evaluation. Application-level evaluations typically involve the assessment of real-world tasks by domain experts, while human-level evaluations require end users to provide feedback on simple tasks. Functional evaluations, on the other hand, rely on the mathematical assessment of accuracies in studies that are not tailored to specific application domains. The interpretability proposed in this project pertains specifically to feature selection and employs an application-level evaluation protocol, which necessitates expert feedback.

Image source: https://www.nature.com/articles/s41929-022-00744-z (Esterhuizen, J.A., Goldsmith, B.R. & Linic, S. Interpretable machine learning for knowledge generation in heterogeneous catalysis. Nat Catal 5, 175–184 (2022).

Relevant Studies of BCI-FeAST Team

Towards Interpretable Feature Selection and Extraction Mechanisms for Asynchronous BCI Systems, TUBITAK-1001, 2024-2027.

Change Point Detection for Time Series

Change Point Detection (CPD) approaches can be employed as a way of finding an adaptive scheme to create the optimal segmentation approach for improving the accuracy of the time series for obtaining the stationarity in classification problem in supervised settings. The figure depicted that the overall accuracy for the classification problem is improved once the CPD applied as a pre-processing mechanism.

Relevant Studies of BCI-FeAST Team

Yıldız, Şeyma, Ballı, Tuğçe, and Yetkin E. Fatih, "Detection of Change Points in a Non-Stationary Time Series via the Fiedler Vectors of the Graph Laplacian", SIAM Applied Linear Algebra, Paris, France, 2024.
Kaçar, Saygın, Ballı, Tuğçe, Yetkin, E. Fatih, "Automatic Segmentation of Time Series Data with PELT Algorithm for Predictive Maintenance in the Flat Steel Industry", UBMK, Antalya, Turkey, 2024.

Asynchronous BCI

Within this BCI-FeAST framework, an efficient and high-accuracy change point detection algorithm will be developed to process raw data continuously and systematically from any asynchronous BCI source (e.g., home care devices, intensive care units, neurorehabilitation applications, etc.).

Relevant Studies of BCI-FeAST Team

Towards Interpretable Feature Selection and Extraction Mechanisms for Asynchronous BCI Systems, TUBITAK-1001, 2024-2027.

Predictive Maintenance

BCI-FeAST also collobarating with industrial organizations (Borçelik) for building interpretable and feasible solutions for predictive maintenance problems.

Relevant Studies of BCI-FeAST Team

Scalable Manifold Learning based on Contour-based Eigenvalue Solvers, TUBITAK-1001, 2021-2024
Recycling Krylov improvement of Non-Linear Dimensionality Reduction Approaches for Large Scale Industrial IoT Data, Bosphore Turkey-France Biliteral Project, 2022-2024.
Özçoban, Kadir, Manguoğlu, Murat and Yetkin, E. Fatih "Investigation of the Effects of Initial Guess and Subspace Iteration Acceleration on Feast Solver", SIAM Applied Linear Algebra, Paris, France, 2024.
Güler, Aykut, Ballı, Tuğçe, and Yetkin E. Fatih, "On Symbolic Prediction of Time Series for Predictive Maintenance Based on SAX-LSTM", UBMK, The Best Paper Award, Antalya, Turkey, 2024.

Privacy Preservation

Modern large-scale production sites are highly data-driven and need large computational power due to the amount of the data collected. Hence, relying only on in-house computing systems for computational workflows is not always feasible. Instead, cloud environments are often preferred due to their ability to provide scalable and on-demand access to extensive computational resources. While cloud-based workflows offer numerous advantages, concerns regarding data privacy remain a significant obstacle to their widespread adoption, particularly in scenarios involving sensitive data and operations. This study aims to develop a computationally efficient privacy protection (PP) approach based on manifold learning and the elementary row operations inspired from the lower-upper (LU) decomposition. This approach seeks to enhance the security of data collected from industrial environments, along with the associated machine learning models, thereby protecting sensitive information from potential threats posed by both external and internal adversaries within the collaborative computing environment.

Relevant Studies of BCI-FeAST Team

CyberMACS-Master’s programme in Applied Cybersecurity, EU-ERASMUS-MUNDUS, 2022-2028
Ballı, Tuğçe, and Yetkin, E. Fatih, "Privacy Preservation for Machine Learning in IIoT Data via Manifold Learning and Elementary Row Operations", submitted to 11th International Conference on Information Systems Security and Privacy.
Hindistan, Yavuz Selim, and E. Fatih Yetkin. "A hybrid approach with GAN and DP for privacy preservation of IIOT data." IEEE Access 11 (2023): 5837-5849.

Page updated

Google Sites

Report abuse