Seminario Machine Learning
Diffusion models for generative artificial intelligence
Ponente: Alberto Suárez (Universidad Autónoma de Madrid)Fecha: viernes 28 de febrero de 2025 - 12:00Lugar: Aula Naranja, ICMAT
Online: https://us02web.zoom.us/j/81739518748?pwd=NAsyeXJj85bpGDzU19KR1xBboSf9YW.1 (ID: 817 3951 8748; password: 622584)
Resumen:
The goal of generative artificial intelligence is to produce samples from a probability distribution whose explicit form is not known. To this end, a set of instances from the unknown distribution is available. In diffusion models, a forward SDE is used to inject noise into the original instances. Then, a neural network is used to model the drift term of a reverse SDE, whose stationary solution approximates the distribution of the original data. One can think of this procedure as learning to invert the arrow of time in a Langevin equation that describes the approach to thermodynamic equilibrium from an initial state with lower entropy. Once the model has been trained, the time-reversed process can be used to create texts, images, or videos with a realistic appearance out of pure noise.
