I'm a 4th-year PhD student supervised by Prof. Pietro Liò and Dr. Ferenc Huszár at the University of Cambridge, Department of Computer Science and Technology, funded by a scholarship from Twitter.
Previously, I spent 5 great years as a Machine Learning Researcher at Bitdefender. I've graduated with valedictorian distinction the BSc in Computer Science from the University of Bucharest in 2016 and the MSc in Artificial Intelligence also from the University of Bucharest in 2018.
My research focuses on Hypergraph Neural Networks. I am particularly interested in developing techniques to better model implicit and explicit higher-order interactions. Additionally, I am interested in exploring the advantages of sheaf theory in machine learning.
Selected Publications
Wasserstein Hypergraph Neural Network
Iulia Duta, Pietro Liò
Preprint
pdf
We are introducing a model that treats the nodes and hyperedge neighbourhood as distributions and aggregate the information using Sliced Wasserstein Pooling. Unlike conventional aggregators such as mean or sum, which only capture first-order statistics, our approach has the ability to preserve geometric properties like the shape and spread of distributions, improving the higher-order modelling.
SPHINX: Structural Prediction using Hypergraph Inference Network
Iulia Duta, Pietro Liò
ICML 2025
pdf
We propose a model that takes advantage of implicit, higher-order interactions for scenarios where these interactions are unobserved or not annotated, enable the application of hypergraph networks on a broader range of applications. SPHINX learns to infer a latent hypergraph structure in an unsupervised way, solely from the final node-level signal.
Sheaf Hypergraph Networks
Iulia Duta, Giulia Cassarà, Fabrizio Silvestri, Pietro Liò
We introduce cellular sheaves for hypergraphs, a mathematical construction that adds extra structure to the conventional hypergraph while maintaining their local, higherorder connectivity. This leads to a novel Sheaf Hypergraph Laplacian that can be used in a machine elrning setup to improve the modelling of higher-order interactions.
Dynamic Regions Graph Neural Networks for Spatio-Temporal Reasoning
Iulia Duta*, Andrei Nicolicioiu*, Marius Leordeanu
We investigate how to create node representations useful for modeling visual interaction using Graph Neural Networks. Our DyReG method can discover salient regions in the scene, without explicit object-level supervision, that correlate with true objects locations. Moreover, it improves the relational processing and obtains superior results on video classification tasks, while being more explainable.
Recurrent Space-time Graph Neural Networks
Andrei Nicolicioiu*, Iulia Duta*, Marius Leordeanu
We propose a neural graph model, recurrent in space and time, suitable for capturing both the local appearance and the complex higher-level interactions of different entities and objects within the changing world scene. Our model is general and could learn to recognize a variety of high level spatio-temporal concepts, obtaining state-of-the-art performance on the challenging Something-Something human-object interaction dataset.
Effective Receptive Field of Graph Neural Networks
Andrei Nicolicioiu*, Iulia Duta*
Technical Report
pdf
We analysed theoretically and empirically the effective receptive field for Graph Convolutional Network and Self-Attention layer.
Mining for meaning: from vision to language through multiple networks consensus
Iulia Duta*, Andrei Nicolicioiu*, Vlad Bogolin, Marius Leordeanu
We design multiple models, to explore different video encoding strategies, to explore intermediate video-language representation and to investigate the gains brought by additional tasks and features. We propose a method for video captioning by selecting from the results of multiple encoder-decoder models. Our method surpassed the state-of-the-art results on the challenging MSR-VTT dataset.
* denotes equal contribution
News
- 06/2025: Invited speaker at NetSCI HONAI Satellite, talking about our recent work SPHINX: Structural Prediction using Hypergraph Inference Network link
- 03/2025: Present a lecture on Hypergraph Representation Learning as part of the L65 Graph Representation Learning Course at University of Cambridge link
- 07/2024: Present a tutorial on Graph Neural Networks as part of the Eastern European Machine Learning Summer School 2024, Novi Sad link
- 09/2023: Give a lecture on Graph Neural Networks at the Mediterranean Machine Learning Summer School 2023, Thessaloniki link
- 12/2022: Co-organise the first Learning on Graph Conference (LOG 2022). link
- 07/2022: Design a practical tutorial on graph neural networks and structural prediction for astern European Machine Learning Summer School 2022 link
Teaching
Representation Learning on Graphs and Networks
University of Cambridge 2022-2025
Teaching Assistant
University of Cambridge 2022-2025
Teaching Assistant
Prepare material and present the practicals/assignments for the course. (MPhil ACS, Part III)
❯
Deep Learning Course
University of Bucharest 2022
Invited speaker
University of Bucharest 2022
Invited speaker
Held a presentation as invited speaker in the MPhil Deep Learning Course (MPhil)
❯
Introduction in Deep Learning
University of Bucharest 2019-2021
Lecturer/Tutor
University of Bucharest 2019-2021
Lecturer/Tutor
Prepare and present the Recurrent Neural Network lecture and mentor several students for their final projects. (MPhil)
❯
Introduction in Graph Nets
University of Bucharest 2021
Lecturer/Tutor
University of Bucharest 2021
Lecturer/Tutor
Prepare and present 2 introductory lectures about Graph Nets, together with Andrei Nicolicioiu. (MPhil)
slides1 video1 slides2 video2
❯
slides1 video1 slides2 video2
Algorithm & Data Structure
University of Bucharest 2017-2019
Teaching Assistant
University of Bucharest 2017-2019
Teaching Assistant
As a teaching assistant, I prepared and presented the laboratories and seminars and helped with the final exam. (1st-year Undergraduate)
❯
Bitdefender Deep Learning Course
Bitdefender 2019
Lecturer/Tutor
Bitdefender 2019
Lecturer/Tutor
Prepare and present the RNNs lecture and offer project mentoship for some collegues. (Industry)
❯