📖 A scientific breakthrough reveals why generative AI learns so well

The meteoric rise of AI has given birth to a new generation of models capable of producing images, sounds or videos of impressive realism. Among them, diffusion models hold a prominent place because by learning from numerous examples, they manage to create content often indistinguishable from real data.

But behind this feat lies a fundamental question: how do these systems manage to invent new data (images, sounds, videos,...), meaning generalize, rather than simply memorizing and then exactly repeating what they have "learned"?


Through an interdisciplinary approach combining statistical physics, computer science and numerical experiments, Tony Bonnaire and his collaborators have made an essential discovery concerning the learning process of diffusion models: they have shed light on two distinct and predictable timescales, with an initial generalization phase independent of the training data, followed much later by a memorization phase dependent on the size of the dataset.

The team shows that the memorization time recedes as the number of training data increases, thus explaining that generative AIs based on diffusion models remain for a long time in a phase where they create new data.

By demonstrating that the observed performance of diffusion models and their practical success rests on a demonstrable and measurable mechanism that naturally delays overfitting, the work of Tony Bonnaire and his collaborators offers a deep and usable understanding of the mechanisms governing modern generative AI.