Enrico Grimaldi

Francesco L. De Faveri introducing our work on “DP-COMET: A Differential Privacy Contextual Obfuscation MEchanism for Texts in NLP”. Co-authored with Guglielmo Faggioli and Nicola Ferro, this research explores how to protect user privacy using contextual obfuscation methods in NLP and IR applications. Abstract: Protecting sensitive information within textual data strongly depends on the context in which the data is presented. However, current privacy-preserving obfuscation mechanisms based on epsilon-Differential Privacy (DP) produce an obfuscated private text, changing the original phrase term-by-term without considering the context in which such a term is placed. This paper introduces DP-COMET, an epsilon-DP obfuscation mechanism that evaluates a text's context before producing its private version. The mechanism defines a representation of the original text that considers the entire context within the text, producing an obfuscated version after adding noise to this representation and depending on the privacy parameter epsilon. We test DP-COMET on different Natural Language Processing (NLP) and Information Retrieval (IR) downstream tasks, and our findings show that our obfuscation mechanism not only achieves comparable performance results to traditional term-by-term mechanisms but also produces obfuscated texts less similar to the originals. To promote the reproducibility of DP-COMET, we make the code publicly available at https://lnkd.in/dcsAWeqx. https://lnkd.in/dvcYP5jp https://lnkd.in/dcsAWeqx

53

Learn how to build an end-to-end reinforcement learning loop to fine-tune an LLM for generating verifiable mathematical proofs. Jacopo Tagliabue shares the full implementation.

18

2 commenti

📢 Highly Cited in #Forecasting! 📖 Forecasting Convective Storms Trajectory and Intensity by Neural Networks ✍️ Niccolò Borghi, Giorgio Guariso and Matteo Sangiorgio 🔗 https://brnw.ch/21wZZkK This study leverages neural networks to improve the prediction of convective storm trajectories and intensity, contributing to more accurate severe weather forecasting and risk mitigation strategies. #WeatherForecasting #NeuralNetworks #SevereWeather #ClimateScience #ArtificialIntelligence

2

Huge news from the lab! We’re heading to #ICLR2026 with 𝐟𝐨𝐮𝐫 accepted papers. Here is a sneak peek at what we’re bringing to the conference: 🍖 Improving Long-Range Interactions in Graph Neural Simulators via Hamiltonian Dynamics by Tai Hoang and collaborators from Davide Bacciu's team at the University of Pisa. We introduce new information-preserving Graph Neural Simulator that enables accurate modeling of long-range interactions in physical systems. 🧌  TROLL: Trust Regions Improve Reinforcement Learning for Large Language Models by Philipp Becker and Niklas Freymuth. We propose a principled trust region projection that greatly improve LLM post-training on math and coding tasks compared to PPO-style clipping. 🧩 Chunking the Critic: A Transformer-based Soft Actor-Critic with N-Step Returns by Dong Tian. Using sequence-aware value estimates lets us capture temporal structure during RL training, which benefits long-horizon and sparse reward tasks. 🚅 Learning Boltzmann Generators via Constrained Mass Transport by our master student Christopher von Klitzing, who was supervised by Denis B. We formalize a new variational framework for generating constrained intermediate distributions for efficient sampling from high-dimensional unnormalized distributions. See you in Brazil! Pre-prints and posters coming soon. 🚀 #MachineLearning #ICLR #DeepLearning

45

How does the order of pretraining data affect neural language models? Our paper, "On the impact of pretraining data ordering in transformer encoder- and decoder-only language models", co-authored with Lucia Domenichelli, Dominique Brunato and Felice Dell'Orletta, has just been published in 💥 Knowledge-Based Systems 💥. In this work, we systematically investigate the role of data ordering during pretraining. We examine multiple curriculum strategies based on sentence length, Gulpease, and Read-It readability scores, along with their inverse variants, and compare them against random data orderings across both encoder and decoder architectures. Our analysis focuses on: • perplexity dynamics • linguistic probing • embedding geometry • downstream tasks performance Our findings indicate that data ordering influences both learning dynamics and the structure of the learned representations. 📄 Paper: https://lnkd.in/dY-Htpeb 💻 Code: https://lnkd.in/dG9VtG_g

69

3 commenti