projects | Roberto Ceraolo

work

CausalQuest - Collecting Natural Causal Questions for AI Agents

We present CausalQuest, a dataset of 13,500 naturally occurring questions sourced from social networks, search engines, and AI assistants. Paper accepted at Causality and Large Models (CALM) workshop at NeurIPS 2024.

OpenRitardi - Open Train Delays

Development of a web app providing visualizations and search functionalities on train delays in Italy. Our goal is to spread awareness about the quality of the service and prompt the company handling the railway system, to publish their data.

Diffusion Models for Video Panoptic Segmentation

Re-implemented of the whole codebase of Pix2Seq-D, a diffusion-based model to approach segmentation tasks, in Pytorch from Tensorflow, and adapted of the architecture to the task of Video Panoptic Segmentation.

Independence regularization for improving out-of-distribution generalization

Study with the goal of improving generalization capabilities of deep learning algorithms. Investigated the usage of Causal Representation Learning to increase the accuracy of a classifier on out-of-distribution data.

US Politicians lexical analysis

Worked with Quotebank, 178 million rows dataset. Using classical Natural Language Processing tools, analyzed the lexicon of American politicians. The goal was to understand whether there are significant differences in terms of complexity and refinement in the use of English by different groups of US politicians.

Synthetic data generation with GANs

Study with the goal to generate realistic synthetic images of C. elegans, in order to apply this architecture in applications where data is scarce, and hard to gather. Used a StyleGAN2-ADA architecture and achieved visually realistic deformations of the real C. elegans images.

Training an AI Assistant for STEM Education

Together with a team of colleagues, developed an AI assistant for STEM students. This included data gathering and preparation, training of a reward model (Deberta), fine-tuning of a large language model (Distilled GPT-2). Used the Antrophic-style Constitutional AI approach in order to bias the model towards clarity, correctness, completeness, and rigour answers.

Robust journey planner with Spark and Hadoop

Leveraging big data from SBB, the Swiss national railway company, built a robust journey planner for the city of Zurich, using Spark and Hadoop. The results obtained were the same as Google Maps’, but we included additional features.