October, Wednesday 11th

14:00 (room 2014, building 660) (see location):

Victor Estrade

(TAU team)

Title: Robust Deep Learning : A case study


We report on an experiment on robust classification against systematic
uncertainties. The literature proposes adversarial and generative
learning, as well as feature construction with auto-encoders.
In both cases, the context is domain-knowledge-free performance.
As a consequence, the robustness quality relies on the representativity
of the training dataset wrt the possible perturbations.
When domain-specific a priori knowledge is available, as in our case,
two specific flavor of DNN are available. One called Pivot Adversarial
Network which uses adversarial training and one called Tangent
Propagation which is a less data-intensive alternative.
We present preliminary experiments and comparison of the two approaches.

Contact: guillaume.charpiat at inria.fr