Seminario

Seminario Machine Learning

Robustness of neural ODEs - a second order adjoint equation approach

Ponente:  Tobias Wöhrer (TU Munich)
Fecha:  lunes 14 de noviembre de 2022 - 12:00
Lugar:  Aula Naranja, ICMAT
Online:  us06web.zoom.us/j/83994891993?pwd=Y29kRnd0YXJGUkZ0OEdrbDJCUG1jQT09

Resumen:

We consider augmented training of ODE based neural networks, such as ResNets, to increase robustness with respect to adversarial attacks. This is done by adding a first order sensitivity term to the loss function, derived from the corresponding robust optimal control problem. To reduce memory cost of the training process, we take an optimize then discretize approach and compute the gradients via solving the adjoint sensitivity equation, which is of second order due to the modified loss function. Based on this approach, we will discuss numerical simulations that provide a better intuition for the robustification process.