MSc Thesis presentation of Dimitrios Krikonis, Wednesday, October 21, 2026
Student: Dimitrios Krikonis
Program: Data Science and Information Technologies / Big Data and
Artificial Intelligence
Title:
Machine Learning for Archaeology: A Human-in-the-Loop Chronology
Prediction System for Pottery from the Athenian Agora
Abstract:
This thesis investigates the use of machine learning as a decision-support
tool for archaeological pottery dating, aiming to provide initial
estimates without replacing expert judgement. First, it conducts a
comparative evaluation of text- and image-processing methods and different
classification and regression algorithms, using 1,910 pottery items from
the The Athenian Agora publication series. Five feature extraction methods
are compared across two modalities: TF-IDF and BERT for text, Canny with
HOG, ResNet and ViT for images, along with their six multimodal
combinations, giving eleven feature sets in total. The results show that
text-based methods outperform image-based approaches, while combining the
two does not provide a reliable improvement, possibly due to the limited
availability of photographs. TF-IDF performs better than BERT on the
dataset, and Vision Transformer (ViT) achieves the strongest results among
image representations. Additionally, a web application was designed and
developed (https://dimitris-kr.github.io/pottery-chronology-predictor/)
that integrates the selected models into a predict–validate–retrain
workflow, allowing archaeologists to review predictions, examine their
uncertainty, validate or correct chronologies, and contribute data for
future model retraining. The system was evaluated by its intended users,
who successfully completed the full workflow. However, whether successive
retraining cycles improve predictive performance remains an open question,
requiring additional validated data collected over time and a dedicated
evaluation methodology.
Date/Time: October 21, 2026 – 13:05 PM EEST
Examination Committee:
Maria Roussou (supervisor), Associate Professor, University of Athens
Alexandros Ntoulas, Assistant Professor, University of Athens
Ioannis Panagakis, Associate Professor, University of Athens
Presentation Link:
https://uoa.webex.com/meet/mariar
MSc Thesis presentation of Eleni Panagiotopoulou Thursday 22/10/2026, 3pm
On Thursday 22/a0/2025 at 3 pm, Eleni Panagiotopoulou, a graduate student in the program “Data Science and Information Technologies, Biomedical Data Science – Bioinformatics” will present their MSc thesis titled:
Actigraphy Feature Representations for Mild Cognitive Impairment Classification: Comparing Summary, Circadian and Wavelet Features across Sensitivity–Specificity Operating Points
Mild cognitive impairment (MCI) is characterized by cognitive deficits that exceed those expected with normal aging without substantially compromising independence in everyday activities. Actigraphy provides a non-invasive means of monitoring rest-activity patterns and may offer additional information for identifying cognitive impairment. This thesis investigated how different actigraphy-derived feature representations distinguish participants with MCI from those with normal cognition.
We compared three feature families: summary statistics, combined cosinor and non-parametric rest-activity descriptors, and discrete wavelet transform (DWT) features. An integrated software pipeline supported signal preprocessing, feature extraction, feature selection, and evaluation of multiple classical machine-learning algorithms within a repeated nested cross-validation framework. We compared configurations that met predefined performance criteria using recall-specificity Pareto fronts and parsimony-based selection. Additional analyses examined age-only models, matched signal lengths, time-of-day windows, model complexity, and hyperparameter tuning.
The feature families supported different recall-specificity trade-offs. Summary statistics achieved the highest median recall, reaching 0.818 at a median specificity of 0.700. Combined cosinor and non-parametric descriptors achieved recall of 0.800 and specificity of 0.750, while DWT features achieved the highest specificity (0.850) with recall of 0.727. Age-only models provided a competitive reference, but incorporating actigraphy features improved the attainable trade-offs. Compact actigraphy representations for wavelet feature extraction retained high specificity, although with lower recall than the leading full-series configuration. Simpler models and configurations without hyperparameter tuning also achieved competitive performance.
These findings indicate that the preferred actigraphy representation depends on the sensitivity-specificity balance sought, rather than a single feature family being universally superior. The thesis contributes an extended analysis pipeline and a systematic comparison of complementary activity representations for MCI classification. The results encourage further investigation of actigraphy as a source of information for clinical decision support, while the limited cohort and exploratory configuration-selection process necessitate independent validation before clinical application.
Examination Committee:
Prof. Em. Elias S. Manolakos (advisor), Dept. of Informatics and Telecommunications, NKUA
Dr. Nikolaos Scarmeas, Professor of Neurology, Department of Medicine, NKUA
Prof. Dimitrios J. Stravopodis, Assoc. Professor of Biology, Department of Biology, NKUA
Meeting Details
Elias Manolakos is inviting you to a scheduled Zoom meeting.
Join Zoom Meeting
https://us02web.zoom.us/j/82515571992?pwd=g0abksV0wrVJmptP5QngbNWQP4DyeR.1
Meeting ID: 825 1557 1992
Passcode: 082617
—
One tap mobile
+302111984488,,82515571992#,,,,*082617# Greece
+302311180599,,82515571992#,,,,*082617# Greece
—
Join by SIP
• 82515571992@zoomcrc.com
Join instructions
https://us02web.zoom.us/meetings/82515571992/invitations?signature=kBwy6M_-PrW9XDPf2ODz5KTSMbQp4zd4TpEBkfH4Ixo
MSc Thesis presentation of Ms. Myrto Potamiti Friday, October 16, 2026
Student: “Myrto Potamiti”
Program: “Data Science and Information Technologies”
Title: “MSTOK: Adapting Discrete Visual Tokenization to Multispectral Earth Observation Imagery”
Abstract:
Recent advances in generative modelling for Earth observation highlight the importance of compact image representations that preserve spatial and spectral information. This thesis presents MSTok, an adaptation of the Open-MAGVIT2 visual tokenizer using lookup-free quantization (LFQ) for 12-band Sentinel-2 Level-2A imagery from TerraMesh. Experiments investigate reflectance preprocessing, discrete code-space capacity, ImageNet-pretrained initialization, architectural adaptations, and joint ImageNet–TerraMesh training. Reconstruction quality is assessed
on 89,088 validation samples at 224×224 resolution using multispectral and RGB perceptual metrics. After one epoch, the standard model trained from scratch outperformed its
ImageNet-initialized counterpart across the reported reconstruction metrics, achieving an FID of 61.38 versus 69.53. A recurring trade-off emerged between multispectral accuracy and RGB perceptual quality: extending linear-adapter training reduced all-band MAE from 0.0480 to 0.0469, while increasing FID from 29.26 to 39.18. High code utilization and accurate standalone adapter reconstruction did not ensure effective end-to-end tokenization. The from-scratch configuration was retained as the preferred overall model within the available training budget.
These findings demonstrate the feasibility of LFQ-based multispectral tokenization and the need to assess multispectral reconstruction fidelity and visible-band perceptual quality jointly.
Date/Time: October 16, 2026 – 14:00 PM.
Examination Committee:
”
Dr. Bill Psomas
Dr. Stavros Perantonis
Dr. Ioannis Kakogeorgiou
“
Presentation link:
https://meet.google.com/dvs-gzfy-nkq
MSc Thesis Presentation by Asimenia Giagli
MSc Thesis Presentation by Asimenia Giagli titled: “A Reproducible Pipeline for Prioritizing Antibody-Accessible Cell-Surface Targets from Single-Cell Transcriptomic Data”
Abstract:
Cell-surface proteins are promising candidates for antibody-based therapies, but their prioritization from transcriptomic data requires more than differential expression alone. This thesis developed a reproducible computational workflow to identify biologically plausible, antibody-accessible surface proteins from single-cell RNA-sequencing data. The workflow was implemented as a modular Nextflow pipeline integrating single-cell preprocessing, cell-type annotation, donor-level pseudobulk differential expression, UniProt protein mapping, transmembrane-topology prediction, GTEx-based normal-tissue filtering, and external validation. It was evaluated using a human cutaneous squamous cell carcinoma dataset containing tumour and matched normal skin samples. Differential expression was performed within annotated cell populations to retain cellular context. Tumour-upregulated genes were mapped to human membrane proteins, filtered for predicted single-pass Type I or Type II topology with sufficient extracellular domains, and screened for low expression in selected vital tissues using GTEx. This progressively reduced the candidate space to a smaller set of cell-type-specific proteins. Validation included comparison with antibody-related IEDB antigens, therapeutic-target annotations, and extracellular-domain BepiPred predictions. The final candidates represent computational priorities rather than validated therapeutic targets and require experimental confirmation of protein expression, surface localization, antibody binding, functional relevance, and normal-tissue safety. Overall, the framework provides a transparent approach for narrowing transcriptomic data toward experimentally testable cell-surface therapeutic candidates.
Date/Time: 16/10/2026 – 13:30 PM
Examination Committee:
Theodore Dalamagas, Research Director, ATHENA Research Center
Dimitrios Stravopodis, Associate Professor, Department of Biology, National and Kapodistrian University of Athens
Eleni Christodoulou, Scientific Associate, ATHENA Research Center
Teams link: https://teams.microsoft.com/meet/362275300616504?p=SX5AkcQxcBxmfu4aOl
UPDATED – PROGRAM FALL SEMESTER 2026-27
Lessons begin on Monday, 5 October (The Python bootcamp will be held the previous week (28/9 – 2/10)
MSc thesis presentation of Vasiliki Koumarela, October 5
Title: “Electricity Price Forecasting with Attention-Based Neural Network Architectures”
Abstract:
In recent years, electricity markets have undergone significant changes, driven by the introduction of auction-based power trading mechanisms, ambitious greenhouse-gas emis-
sion targets, and the increasing penetration of Renewable Energy Sources (RES), which together make electricity price forecasting a highly volatile, multifactorial problem. These developments have heightened the importance of quantifying forecast uncertainty and have motivated growing interest in probabilistic forecasting, particularly with deep learning models. This thesis contributes to probabilistic electricity price forecasting by applying and evaluating modern attention-based architectures, namely the Transformer, the Temporal Fusion Transformer (TFT), PatchTST, and the foundation model Chronos-2, adapted to produce quantile forecasts. The models are first evaluated on the GEFCom2014 dataset under the same experimental setting as the original competition, on which, to the best of our knowledge, such deep learning architectures have not previously been applied; their results are compared against both classical statistical and neural benchmarks, as well as the competition-winning solutions. To assess robustness under more realistic and demanding conditions the same framework is then applied to recent Greek electricity market data, which exhibit high volatility, missing values, and negative prices. The forecasting task is to predict the next 24 hours from a historical observation window together with available exogenous features, with performance evaluated using the pinball loss alongside reliability and sharpness metrics. The results show that the proposed models consistently outperform the classical benchmarks on GEFCom2014, in some cases matching or exceeding the competition-winning solutions. On the more complex Greek data the models remain effective, though with some degradation attributable to the increased difficulty of the series, indicating room for further improvement. Overall, Chronos-2 achieves the strongest performance, followed closely by the Transformer.
Date-Time: 5/10/2026 – 16:00 PM
Examination Committee:
Associate Professor Aggelos Pikrakis, University of Piraeus
Dr. Vassilis Katsouros, Director, ILSP/Athena Research Center
Dr. Kosmas Kritsis, ILSP/Athena Research Center
Msc Thesis Presentation
Monday, October 5 · 16:00pm – 17:00pm
Time zone: Europe/Athens
Webex joining info
Video call link: https://unipi.webex.com/meet/webex-host2