Seminario impartdo por el Dr. Pietro Rotondo. Università degli Studi di Milano Statale.
Resumen: Modern deep neural networks (DNNs) represent a formidable challenge for theorists: according to the commonly accepted probabilistic framework that describes their performance, these architectures should overfit due to the huge number of parameters to train, but in practice they do not. Here we employ results from replica mean field theory to compute the generalisation gap of machine learning models with quenched features, in the teacher-student scenario and for regression problems with quadratic loss function. Notably, this framework includes the case of DNNs where the last layer is optimised given a specific realisation of the remaining weights. We show how these results – combined with ideas from statistical learning theory – provide a stringent asymptotic upper bound on the generalisation gap of fully trained DNN as a function of the size of the dataset P. In particular, in the limit of large P and Nout (where Nout is the size of the last layer) and Nout ≪ P , the generalisation gap approaches zero faster than 2Nout/P, for any choice of both architecture and teacher function. Notably, this result greatly improves existing bounds from statistical learning theory. We test our predictions on a broad range of architectures, from toy fully-connected neural networks with few hidden layers to state-of-the-art deep convolutional neural networks.
- Fecha: Miércoles 8 de junio de 2022
- Lugar: Aula de Física Computacional. Facultad de Ciencias.
- Horario: 11:00 h
- Organiza: Grupo de Física Estadística y de los Sistemas Complejos (FQM165)
- Más información:
Dr. Carlos Pérez-Espigares
Profesor Contratado Doctor
Depto. de Electromagnetismo y Física de la Materia
Instituto Carlos I de Física Teórica y Computacional
Universidad de Granada
Avenida Fuentenueva, 18071 Granada (España)
Tel: +34 958 243385
e-mail: carlosperez@ugr.es