Projects
With the increasing use of personal images on the internet and social platforms, privacy concerns have grown significantly. Traditional methods for protecting privacy, such as blurring or pixelating faces, are no longer sufficient to prevent the identification of individuals.
Deep learning has become one of the best approaches to addressing tasks that are repetitive and based on trainable knowledge. The main idea is to apply it to the eHealth domain to perform a guided tasks by simulating the behavior of a physician. Goal
Theoretical and critical analysis of Transformer architecture; Development of new Transformer architectures including new Layers (Gated Residual Kolmogorov-Arnold Networks, Feature Pyramid Network), Positional Encoders (Attention-Leveraged Rotary Positional Embeddings), Attentions (Multi-scale attention, Attention with uncertainty estimation) Implementation of the new architectures in an LLM and performance evaluation by experimental testing; Statistical analysis of the obtained performance and selection of the best architectures; Development of a framework for optimal selection of architectures and parameters.
Among applications of LLMs, in this research area we investigate those related to problems in Society, such as gender-based violence and discrimination, or the dissemination of propaganda messages.
A small language model (SLM) is a machine learning model typically based on a large language model (LLM) but of greatly reduced size. An SLM retains much of the functionality of the LLM from which it is built but with far less complexity and computing resource demand.
Is there any trade-off between performance of recommendation algorithms and their carbon footprint?
In recent years, our focus has been on addressing the critical issue of replicability in research papers about Recommender Systems. This is a key area of concern, as it directly impacts the reliability and trustworthiness of our findings. Our goal is to develop a novel RecSys architecture, including those based on LLMs, that can be replicated with consistent results. If this is not possible, we are committed to releasing a standard pipeline that can be used to reproduce our results.
Food recommender systems aim to identify suitable recipes for a target users.
Multimodal approaches leveraging information from diverse sources are becoming more and more common. Bridging the gap between these diverse worlds is becoming more and more essential.
LaikaLLM is a software, for researchers, that helps in setting up a repeatable, reproducible, replicable protocol for training and evaluating multitask LLM for recommendation!
A knowledge graph represents a network of real-world entities such as objects, events, situations, or concepts and illustrates the relationship between them. Integrating such structured information into an LLM’s knowledge is not easy.
How to exploit and adapt LLMs for recommendation tasks (fine-tuning, knowledge injection, direct prompting, etc.) How to exploit and adapt LMMs for recommendation tasks (input=text? image? output=text and image? only image? AI-generated image or video? many possible strategies) How to exploit the information encoded in KGs for recommendation? (architectures, encoding, information represented in the graph, pre-training vs. direct learning, etc.)
The JARVIS project involve the use of LLM with a focus on personalization. Indeed, the idea is to investigate the possibility of designing an Digital Agent able to use Adaptive Personalization. The AI needs to continuously learn from the user’s inputs and feedback to improve and adjust suggestions, making the experience feel tailored and unique. LLMs could adjust their tone, style, and even choice of content based on a user’s past interactions. For instance, they might recognize if a user prefers concise responses, casual language, or specific domains of knowledge (e.g., a technical user might prefer more detail in scientific explanations). Models could incorporate memory modules (Memory Mechanisms) that help them retain important context from previous interactions, enabling more continuous and coherent conversations. However, memory retention must be handled carefully to avoid storing unnecessary or sensitive information. Research could focus on enabling LLMs to create more sophisticated user models that evolve as user preferences change over time, dynamically updating without needing explicit user input or resetting preferences (Long-Term User Models).
Large Language Models are revolutionizing how users access information. Public Administration is characterized by several documents, laws, regulations (which are not easy to find), and a specific domain lexicon.
Conversational Recommender Systems provide suggestions based on multi-turn dialogues between a chatbot and a user, where preferences and needs are expressed by using natural language statements.
As the Large Language Model (LLM) field accelerates in its breakthroughs, our Anita model must keep pace to maintain its edge. Daily, researchers introduce innovative methods to refine, reason, nudge, and adjust these intelligent tools. To stay at the forefront, we must not only keep up with the latest advancements, but also deliberately direct their application to enhance our Anita model, infusing it with novel capabilities, tailoring its expertise to specific areas of interest, and, most importantly, ensuring its Italian character remains a defining trait.”