[01017] Feature Collisions in Neural Networks: Theory and Practice
Session Time & Room : 3D (Aug.23, 15:30-17:10) @E502
Type : Contributed Talk
Abstract : Deep neural networks are behind many breakthroughs in the last decade, but much of their behavior remains poorly understood. In particular, under some conditions, neural networks can be insensitive to changes of large magnitude, in which case the features are said to collide. We will discuss necessary conditions for such feature collisions to occur, and we will introduce the null-space method, a numerical approach to create data points with colliding features for many vision tasks.
