Iason Skylitsis - AI Master's Student | University of Amsterdam

Iason Skylitsis

Mail | LinkedIn | Google Scholar | GitHub | CV

Last updated: March 2026

About Me

AI Master's student at the University of Amsterdam, currently conducting my thesis at QurAI, Amsterdam UMC, on few- and zero-shot learning for automated body composition analysis in CT scans, developing label-efficient segmentation frameworks using foundation model encoders and prototypical networks, under supervision of Dr. Dimitrios Karkalousos.

Previously I researched causal bandits, under supervision of PhD candidate Roel Hulsman and Assistant Professor Sara Magliacane, developing frameworks that can learn optimal interventions when both the causal structure is unknown and some variables cannot be directly manipulated.

With a background in Computer Science, I bring a hands-on approach to machine learning research, bridging theoretical advances with practical implementations. My industry experience as a software engineer at Twiki Solutions included leading data-driven services and deploying an ETA prediction model for production logistics systems.

My research interests span causality, representation learning, and applied ML. I'm particularly drawn to problems that require both theoretical rigor and practical implementation—developing methods that advance our understanding while solving meaningful challenges.

Research Projects

Few- and Zero-Shot Learning for Automated Body Composition Analysis in CT Scans Iason Skylitsis, Dimitrios Karkalousos Ongoing Thesis Research Developing label-efficient segmentation frameworks for automated body composition analysis in CT scans Leveraging foundation model encoders and prototypical networks for few- and zero-shot learning Addressing the need for reduced annotation burden in medical image segmentation
Towards Structural Causal Bandits with Non-Manipulable Variables and Unknown Causal Structure Iason Skylitsis, Roel Hulsman, Sara Magliacane Links: Paper \| Code Extending the Structural Causal Bandits framework to operate on equivalence classes (CPDAGs, PAGs) learned through causal discovery Incorporating non-manipulable variables (Lee & Bareinboim 2019) into the intervention-selection process to handle realistic decision-making constraints
Injecting Image Guidance into Text-Conditioned Diffusion Models at Inference Agata Żywot, Iason Skylitsis, Thijmen Nijdam, Zoe Tzifa-Kratira, Derck W. E. Prinzhorn, Konrad Szewczyk, Aritra Bhowmik Links: Paper \| Code \| Poster Introduced Visual Concept Fusion (VCF), the first method enabling dual conditioning on both image and text prompts at inference time without retraining Developed a lightweight aligner using InfoNCE and cross-attention reconstruction losses to map image tokens to text embedding space Implemented a fusion strategy that preserves both textual and visual semantics Added optional Prompt-Noise Optimization module for test-time refinement
(Even More) Efficient Equivariant Transfer Learning from Pretrained Models Mikhail Vlasenko, Ádám Divák, Iason Skylitsis, Milan Miletić, Zoe Tzifa-Kratira Links: Blogpost \| Code Reproduced and extended the original paper by Basu et al. (2023) Verified results across missing experiments and expanded the discussion on weight patterns learned by λ‑equitune Proposed and implemented equiattention, an extension utilizing an Attention layer to potentially improve performance over λ‑equitune
Reproduction and Extension of "Robust Fair Clustering: A Novel Fairness Attack and Defense Framework" Iason Skylitsis, Zheng Feng, Idries Nasim, Camille Niessink Transactions on Machine Learning Research (TMLR), 2024 Also: NeurIPS 2024 (ML Reproducibility Challenge poster) Links: Paper \| Code \| Poster \| Slides Reproduced and extended the ICLR'23 paper by Chhabra et al., refining the codebase Introduced additional metrics and datasets for comprehensive evaluation Explored novel attack strategies including targeted attacks Conducted ablation studies to validate original claims and broaden the research landscape
Evaluation of Supervision in Deep Learning Networks for Accelerated-MRI Reconstruction Iason Skylitsis, Dimitrios Karkalousos, Matthan W. A. Caan Links: Paper \| Code \| Slides Investigated the efficiency and reconstruction quality of different supervised and self-supervised methods for MRI reconstruction Extensively compared pre-trained CIRIM to various methods including Noise2Recon, SSDU, U-Net, ResNet, and traditional Compressed Sensing approaches Evaluated trade-offs between training efficiency and reconstruction quality
Deep Learning and Explainable Artificial Intelligence Techniques for Alzheimer's Disease Detection on MRI Iason Skylitsis, Athanasios Voulodimos, Eftychios Protopapadakis Links: Paper Developed an Alzheimer's disease detection framework using deep learning and transfer learning on 2180 MRI images from the ADNI dataset Implemented a comprehensive preprocessing pipeline for MRI data Improved model explainability using Grad‑CAM visualization techniques

Source: adapted from Thijmen Nijdam's fork of Alexander Timans' fork of Dharmesh Tailor's fork of Leonid Keselman's fork of John Barron's website.