
Smart equipment and continuous monitoring for water network
RESEARCH PROJECTS
Development and definition of the neural architecture of the language model with associative memory mechanisms, and its experimentation
Current LLMs learn through statistical accumulation over vast amounts of data within largely “black box” architectures: the process by which information is learned remains opaque and difficult to interpret. This creates a critical trade-off: generative power versus controllability and transparency.
Developed in partnership with the University of Rome Tor Vergata and funded by the Italian Ministry of University and Research’s Fund for Applied Sciences, MeMo is a cutting-edge research project proposing a fundamental shift: moving from models that implicitly learn what to remember to models that make memory mechanisms structured, inspectable, and editable.
The innovation lies in a revolutionary architecture based on Correlation Matrix Memories, explicit associative memories materializing memory as a queryable data structure rather than distributing it across the network’s parameters.
We contributed to the development and definition of the neural architecture of the new language model, as well as to its experimentation, with the aim of achieving greater control over the information embedded in the model and its reliability.
In addition, we provide support for the experimentation and evaluation of this new class of models, leveraging the know-how gained from training Velvet models in terms of:
related to computation and lower training costs
of specific information
regarding how and why responses are generated
to ensure trustworthy, governable, and regulation-compliant AI