Thomas Melistas
PhD student (NKUA) & Academic fellow
National and Kapodistrian University of Athens
SHORT BIO
Thomas holds an MEng in Electrical and Computer Engineering from the National Technical University of Athens, with specialization in Computer Science. The topic of his thesis was Lyrics and Vocal Melody Generation conditioned on Accompaniment.
He has extensive work experience as Machine Learning/MLOps Engineer and has published papers in various conferences on the fields of computational creativity, music information retrieval, affective computing and speech processing.
He has showcased great interest in the application of AI in the arts and especially music. He has developed various related projects and has participated in workshops and hackathons, being awarded the Best Hack / Research Direction Award in HAMR 2020.
His interests include music, photography and poetry among others.
Thomas’ research interests include multimodal and unsupervised learning, interpretability and causality.
RESEARCH INTERESTS
Thomas’ research interests include multimodal and unsupervised learning, interpretability and causality.
Specifically, his research will focus on areas such as causal representation learning and generative models, as well as their biomedical applications.
PhD research on "Benchmarking Counterfactual Image Generation"
Abstract: Generative AI enables powerful visual editing, but realistic counterfactual image generation demands causal consistency and lacks clear ground truth, complicating evaluation. This work introduces a unified benchmarking framework, integrating existing methods, datasets, and causal graphs. Results highlight Hierarchical VAEs’ superiority, with a user-friendly, extensible Python package for future community research.
⚬ Recent publication: UCRL@ICLR 2026 paper "Identifiable Estimation of Causal Concept Effects under Visual Latent Confounding"
PhD research on "Benchmarking Counterfactual Image Generation"
Abstract: Generative AI enables powerful visual editing, but realistic counterfactual image generation demands causal consistency and lacks clear ground truth, complicating evaluation. This work introduces a unified benchmarking framework, integrating existing methods, datasets, and causal graphs. Results highlight Hierarchical VAEs’ superiority, with a user-friendly, extensible Python package for future community research.
⚬ Recent publication: UCRL@ICLR 2026 paper "Identifiable Estimation of Causal Concept Effects under Visual Latent Confounding"
