Βραβείο καλύτερης δημοσίευσης σε Περιοδικό

• • Apostolos Evangelidis – Efficient deep Q-learning for industrial equipment calibration in elevator manufacturing

 
Σχόλια επιτροπής αξιολόγησης:

Originality – Excellent originality, devising novel mathematical formulation and subsequently applying to real manufacturing facilities for extensive testing.

Importance – Real world application and real world industrial setting for experimentation.

Rigour – Sound and rigorous methodological approach, supported by extensive experimentation
 

Βραβείο καλύτερης δημοσίευσης σε Συνέδριο

• Christos Koutlis – Leveraging representations from intermediate encoder-blocks for synthetic image detection

 
Σχόλια επιτροπής αξιολόγησης:

Originality – The paper introduces a novel approach to synthetic image detection by leveraging intermediate representations from CLIP’s image encoder. This is an interesting contribution to the field, since prior work mostly focused on final-layer embeddings. The design of the Trainable Importance Estimator adds novelty in the architecture.

Importance – With the exponential rise of generative AI, detecting synthetic media is a real, pressing technical challenge with important societal implications. The method achieves strong generalization across a broad range of generative models, making it a good candidate for real-world deployment in relevant systems, e.g. trust and safety ones. The algorithm outperforms state-of-the-art methods by +10.6% accuracy and trains very fast. It is important to note that the paper has already cited 23 times according to Google Scholar

Rigour – The experimental setup is extensive, covering 20 datasets and including comparisons with state-of-the-art methods, ablation studies, and robustness tests. The model is evaluated in multiple training configurations with clear metric reporting. The level of empirical detail is very good.
 

Βραβείο καλύτερης δημοσίευσης φοιτητή σε Περιοδικό

• Thanasis Kotsiopoulos – Revolutionizing defect recognition in hard metal industry through AI explainability, human-in-the-loop approaches and cognitive mechanisms

 
Σχόλια επιτροπής αξιολόγησης:

Originality – This paper presents an innovative approach to defect recognition in the hard metal industry by integrating explainable AI (XAI), human-in-the-loop (HITL) techniques, and cognitive retraining mechanisms. Unlike conventional automated defect detection systems that operate as black-box models, this study emphasizes AI transparency and human-AI collaboration. The inclusion of interpretable AI models ensures that AI-driven decisions are understandable to operators, allowing informed interventions and refinements.

Importance – The paper addresses key challenges in industrial defect detection, particularly trust and adaptability in AI systems. Traditional AI-based quality control methods often struggle with operator acceptance and reliability concerns, as their predictions lack intuitive explanations. By incorporating XAI, the platform provides insightful justifications for each AI decision. Furthermore, HITL mechanisms allow real-time expert feedback, enabling models to learn from human corrections and continuously improve their detection accuracy.

Rigour – The research presents a structured methodology, detailing model architecture, image acquisition techniques, and micro-service-based system design. The introduction of retraining mechanisms ensures AI models remain robust over time. The study uses machine and deep learning algorithms for defect classification and localization, leveraging industry-standard AI techniques. However, while simulation results are promising, further experimental validation in real data would be much desirable. Additionally, conducting a comparative analysis against existing AI-based defect detection solutions would strengthen the paper.
 

Βραβείο καλύτερης δημοσίευσης φοιτητή σε Συνέδριο

• Anestis Kastellos – FedHARM: Harmonizing Model Architectural Diversity in Federated Learning

 
Σχόλια επιτροπής αξιολόγησης:

Originality – Unlike standard FL approaches that require homogeneous models for weight aggregation, FedHARM focuses on harmonizing representations rather than model parameters through a hybrid training method.

Importance – The architecture heterogeneity problem appears to be a novel problem that this work handles in FL. As such it is an important contribution to the community although experimentation was performed in simple datasets.

Rigour – The methodology is clearly described, includes implementation details, and presents quantitative evaluations. The experiments are thorough across multiple datasets, architectures, and client counts. The question is how and whether the methodology adapts to other architectures, and different and more modern datasets.