I am a fifth year PhD student in Computer Science at Harvard University, advised by Demba Ba. My research interests revolve around deep learning, mathematics, and optimization. I'm particularly interested in representations, geometry and topology, and dynamics in deep learning. During the summer of 2021 I interned at Amazon Web Services where I had the pleasure to be hosted by Karim Helwani. I finished my undergraduate studies at the National Technical University of Athens in Greece, where I conducted my thesis in tropical geometry under the supervision of Petros Maragos.
Mentoring: As a first-generation student, I'm passionate about providing help and guidance to people who have no access to it. This led to the creation of a mentoring initiative, MentoRes, along with Konstantinos Kallas, that intends to help people applying for PhD programs in Computer Science or engineering fields. I'm always happy to chat and give advice, even if it's unrelated to PhD applications.
Personal: I was born and spent the first couple of years of my life in Crete. I then moved to Athens, where I spent most of my life before coming to the US for my PhD. I enjoy quotes a lot and keep lists of my current favorites. My favorite book is "The Cider House Rules" by John Irving, I'm a movie buff with a soft spot for tasteful thrillers and horror movies, and I enjoy skill-based video games. I play board games weekly, mainly social deception and minmax euro games. I love national parks, but unfortunately my quest to visit all of them has lagged behind due to COVID.
Neural networks have achieved state-of-the-art results in many fields, however it's not clear how to design networks the respect, and account for, the inductive biases of a particular task. This line of work studies principled ways to encode, and learn, inductive biases in neural networks, resulting in better performance, interpretability, and reduced numbers of parameters.
- "Learning linear groups in neural networks"
Learning meaningful representations is instrumental for most learning problems. Representations with appropriate properties can reduce model sizes, improve performance. and be more interpretable by being better suited to the task at hand. This line of work studies how to construct networks whose representations satisfy these given constraints.
Linear algebra powers most learning systems today; however, there are nonlinear algebras that are more expressive and better suited for certain applications. This line of work studies how these algebras can be used in various machine learning settings to model problems of interest, resulting in new algorithms, unified formulations, and intuitive explanations.
- "Tropical Geometry and Machine Learning"
- "Tropical modeling of weighted transducers algorithms on graphs"