Research Explorer

Project

A formal characterization of the robustness of deep learning techniques
Code
01D31417
Duration
01 October 2017 → 30 September 2018
Funding
Regional and community funding: Special Research Fund
Promotor
Research disciplines
  • Natural sciences
    • Bioinformatics and computational biology not elsewhere classified
Keywords
recurrent neural networks machine learning deep learning convolutional neural networks robustness
 
Project description
 

Deep neural networks appear to be very sensitive to small perturbations to their input that should leave their output unchanged. However, specially crafted tiny perturbations can change the output of these networks dramatically, a phenomenon which is not yet well understood. In this work, we will aim to find a formal characterization of these perturbations.