ITU Vision & AI Lab

The ITU Vision & AI Lab conducts research at the intersection of machine learning, computer vision, and probabilistic artificial intelligence. Our current research focuses on representation learning, multimodal and latent-variable models, uncertainty-aware learning, self-supervised learning, and AI methods for scientific and biomedical applications.

The lab is directed by Prof. Dr. Gozde Unal at the Faculty of Computer and Informatics Engineering, Istanbul Technical University.

news

June 2026	New preprint: Neural Conjugate Aggregation: Identifiable Unsupervised Multi-Sensor Regression under Heterogeneous Sensor Bias is now available on arXiv. The work introduces NCAM, a probabilistic framework for unsupervised multi-sensor fusion with uncertainty quantification and conformal prediction.
June 2026	New preprint: Masked and Predictive Self-Supervised Foundation Models for 3D Brain MRI is now available on arXiv. The study investigates masked reconstruction and predictive representation learning for self-supervised pretraining in 3D brain MRI.
January 2026	New publication: Disentanglement with Factor Quantized Variational Autoencoders by Gulcin Baykal, Melih Kandemir, and Gozde Unal was published in Neurocomputing.
September 2025	Prof. Gozde Unal returned to ITU following her sabbatical leave and resumed her research activities at the ITU Faculty of Computer and Informatics Engineering.
2025	New paper: UniMLR: Modeling Implicit Class Significance for Multi-Label Ranking presented at ECAI 2025.
2024–2025	Prof. Gozde Unal spent her sabbatical at NYU Langone Health as a Senior Research Scientist, working on self-supervised representation learning, MRI foundation models, and uncertainty-aware machine learning.
December 2024	Congratulations to Dr. Gulcin Baykal Can on successfully completing her PhD! Her doctoral research focused on discrete representation learning and generative models.

Selected Publications

2026

CAADRIA

Integrating Perceptual and Computational Frameworks for Walkability Assessment in Campus Environments Using AI-Based Models

Sena Kaynarkaya , Aslı Çekmiş , İsmail Çetin , and 2 more authors

In Proceedings of the 31st International Conference on Computer-Aided Architectural Design Research in Asia , 2026

Abs Bib HTML

@inproceedings{kaynarkaya2026campus,
  title = {Integrating Perceptual and Computational Frameworks for Walkability Assessment in Campus Environments Using AI-Based Models},
  author = {Kaynarkaya, Sena and Çekmiş, Aslı and Çetin, İsmail and Şahin, Yusuf Hüseyin and Ünal, Gözde},
  booktitle = {Proceedings of the 31st International Conference on Computer-Aided Architectural Design Research in Asia},
  volume = {3},
  pages = {543--552},
  year = {2026},
  url = {https://papers.cumincad.org/data/works/att/caadria2026_208.pdf},
}

arXiv

Neural Conjugate Aggregation: Identifiable Unsupervised Multi-Sensor Regression under Heterogeneous Sensor Bias

Muhammed Faruk Aytin , Zehra Demir , Alper Ünal , and 2 more authors

2026

Abs Bib HTML

@article{aytin2026neural,
  title = {Neural Conjugate Aggregation: Identifiable Unsupervised Multi-Sensor Regression under Heterogeneous Sensor Bias},
  author = {Aytin, Muhammed Faruk and Demir, Zehra and Ünal, Alper and Marshall, Julian and Ünal, Gözde},
  booktitle = {arXiv Preprint},
  year = {2026},
  url = {https://arxiv.org/abs/2606.22200},
}

arXiv

Masked and Predictive Self-Supervised Foundation Models for 3D Brain MRI

Esra Ergün , Hersh Chandarana , Dan Sodickson , and 1 more author

2026

Abs Bib HTML

@article{ergun2026masked,
  title = {Masked and Predictive Self-Supervised Foundation Models for {3D} Brain {MRI}},
  author = {Ergün, Esra and Chandarana, Hersh and Sodickson, Dan and Ünal, Gözde},
  booktitle = {arXiv Preprint},
  year = {2026},
  url = {https://arxiv.org/abs/2606.13315},
}

IEEE TMI

BONBID-HIE 2023: Lesion Segmentation Challenge in BOston Neonatal Brain Injury Data for Hypoxic Ischemic Encephalopathy

Rina Bao , Anna N. Foster , Ya’Nan Song , and 28 more authors

IEEE Transactions on Medical Imaging, 2026

Abs Bib HTML

@article{bao2026bonbidhie,
  title = {{BONBID-HIE} 2023: Lesion Segmentation Challenge in {BOston} Neonatal Brain Injury Data for Hypoxic Ischemic Encephalopathy},
  author = {Bao, Rina and Foster, Anna N. and Song, Ya'Nan and Vyas, Rutvi and Kesri, Ankush and Toubal, Imad Eddine and Kazemi, Elham Soltani and Rahmon, Gani and Kucukpinar, Taci and Almansour, Mohamed and Ho, Mai-Lan and Palaniappan, K. and Ninalga, Dean and Koirala, Chiranjeewee Prasad and Mohapatra, Sovesh and Schlaug, Gottfried and Wodzinski, Marek and Muller, Henning and Ellis, David G. and Aizenberg, Michele R. and Aydın, M. Arda and Abdinli, Elvin and Unal, Gozde and Tahmasebi, Nazanin and Punithakumar, Kumaradevan and Song, Tian and Peng, Yun and Bates, Sara V. and Hirschtick, Randy and Grant, P. Ellen and Ou, Yangming},
  journal = {IEEE Transactions on Medical Imaging},
  volume = {45},
  number = {4},
  pages = {1711--1725},
  year = {2026},
  doi = {10.1109/TMI.2025.3638977},
  url = {https://doi.org/10.1109/TMI.2025.3638977},
}

Neurocomputing

Disentanglement with Factor Quantized Variational Autoencoders

Gulcin Baykal , Melih Kandemir , and Gozde Unal

Neurocomputing, 2026

Abs Bib HTML

@article{baykal2026factorqvae,
  title = {Disentanglement with Factor Quantized Variational Autoencoders},
  author = {Baykal, Gulcin and Kandemir, Melih and Unal, Gozde},
  journal = {Neurocomputing},
  volume = {661},
  pages = {131968},
  year = {2026},
  doi = {10.1016/j.neucom.2025.131968},
  url = {https://doi.org/10.1016/j.neucom.2025.131968},
}

2025

ECAI

UniMLR: Modeling Implicit Class Significance for Multi-Label Ranking

V. Bugra Yesilkaynak , Emine Dari , Alican Mertan , and 1 more author

In Proceedings of the European Conference on Artificial Intelligence , 2025

Abs Bib HTML

@inproceedings{yesilkaynak2025unimlr,
  title = {{UniMLR}: Modeling Implicit Class Significance for Multi-Label Ranking},
  author = {Yesilkaynak, V. Bugra and Dari, Emine and Mertan, Alican and Unal, Gozde},
  booktitle = {Proceedings of the European Conference on Artificial Intelligence},
  pages = {3242--3249},
  year = {2025},
  doi = {10.3233/FAIA251191},
  url = {https://doi.org/10.3233/FAIA251191},
}

CVPRW

ITACLIP: Boosting Training-Free Semantic Segmentation with Image, Text, and Architectural Enhancements

M. Arda Aydın , Efe Mert Çırpar , Elvin Abdinli , and 2 more authors

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , 2025

Abs Bib HTML

@inproceedings{aydin2025itaclip,
  title = {{ITACLIP}: Boosting Training-Free Semantic Segmentation with Image, Text, and Architectural Enhancements},
  author = {Aydın, M. Arda and Çırpar, Efe Mert and Abdinli, Elvin and Unal, Gozde and Sahin, Yusuf H.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
  pages = {4181--4191},
  year = {2025},
  doi = {10.1109/CVPRW67362.2025.00398},
  url = {https://openaccess.thecvf.com/content/CVPR2025W/PixFoundation/html/Aydin_ITACLIP_Boosting_Training-Free_Semantic_Segmentation_with_Image_Text_and_Architectural_CVPRW_2025_paper.html},
}

2024

PR
EdVAE: Mitigating Codebook Collapse with Evidential Discrete Variational Autoencoders

Gulcin Baykal , Melih Kandemir , and Gozde Unal

Pattern Recognition, 2024

Abs Bib HTML

Codebook collapse is a common problem in training deep generative models with discrete representation spaces like Vector Quantized Variational Autoencoders (VQ-VAEs). We observe that the same problem arises for the alternatively designed discrete variational autoencoders (dVAEs) whose encoder directly learns a distribution over the codebook embeddings to represent the data. We hypothesize that using the softmax function to obtain a probability distribution causes the codebook collapse by assigning overconfident probabilities to the best matching codebook elements. In this paper, we propose a novel way to incorporate evidential deep learning (EDL) instead of softmax to combat the codebook collapse problem of dVAE. We evidentially monitor the significance of attaining the probability distribution over the codebook embeddings, in contrast to softmax usage. Our experiments using various datasets show that our model, called EdVAE, mitigates codebook collapse while improving the reconstruction performance, and enhances the codebook usage compared to dVAE and VQ-VAE based models.
@article{baykal2023edvae, title = {EdVAE: Mitigating Codebook Collapse with Evidential Discrete Variational Autoencoders}, author = {Baykal, Gulcin and Kandemir, Melih and Unal, Gozde}, journal = {Pattern Recognition}, volume = {156}, pages = {110792}, year = {2024}, doi = {10.1016/j.patcog.2024.110792}, url = {https://doi.org/10.1016/j.patcog.2024.110792}, }

AUTCON

Semi-Automated Minimization of Brick-Mortar Segmentation Errors in 3D Historical Wall Reconstruction

Mustafa Cem Güneş , Alican Mertan , Yusuf H. Sahin , and 2 more authors

Automation in Construction, 2024

Abs Bib HTML

@article{gunes2024brickmortar,
  title = {Semi-Automated Minimization of Brick-Mortar Segmentation Errors in {3D} Historical Wall Reconstruction},
  author = {Güneş, Mustafa Cem and Mertan, Alican and Sahin, Yusuf H. and Unal, Gozde and Özkar, Mine},
  journal = {Automation in Construction},
  volume = {167},
  pages = {105693},
  year = {2024},
  doi = {10.1016/j.autcon.2024.105693},
  url = {https://doi.org/10.1016/j.autcon.2024.105693},
}

2022

NeuRIPS

How to combine variational bayesian networks in federated learning

Atahan Ozer , Kadir Burak Buldu , Abdullah Akgül , and 1 more author

2022

Abs Bib HTML

@article{ozer2022combine,
  title = {How to combine variational bayesian networks in federated learning},
  author = {Ozer, Atahan and Buldu, Kadir Burak and Akg{\"u}l, Abdullah and Unal, Gozde},
  booktitle = {International Workshop on Federated Learning: Recent Advances and New Challenges in Conjunction with NeurIPS 2022 (FL-NeurIPS'22)},
  year = {2022},
  url = {https://arxiv.org/abs/2206.10897},
}

ICLR

Evidential turing processes

Melih Kandemir , Abdullah Akgül , Manuel Haussmann , and 1 more author

2022

Abs Bib HTML

@article{kandemir2021evidential,
  title = {Evidential turing processes},
  author = {Kandemir, Melih and Akg{\"u}l, Abdullah and Haussmann, Manuel and Unal, Gozde},
  booktitle = {International Conference on Learning Representations},
  year = {2022},
  url = {https://arxiv.org/abs/2106.01216},
}