Correcting Deep Neural Networks with Small, Generalizing Patches

Sotoudeh, Matthew and Thakur, Aditya V.
NeurIPS 2019 Workshop on Safety and Robustness in Decision Making, 2019

We consider the problem of patching a deep neural network: applying a small change to the network weights in order to produce a desired change in the clas- sifications made by the network. We motivate this problem using ACAS Xu, a well-studied neural network intended to act as an aircraft collision-avoidance system. Our technique works over infinite patching regions, is based on an SMT formulation of the problem, and has a number of desirable convergence properties. To make this approach efficient, we introduce a symbolic representation of neural networks and a generalization of ReLU neural networks (Masking Networks). We show that our approach can produce highly effective and generalizing patches with a very small number of weight changes.

PDF    

@inproceedings{SRDM2019,
  author = {Sotoudeh, Matthew and Thakur, Aditya V.},
  title = {Correcting Deep Neural Networks with Small, Generalizing Patches},
  booktitle = {NeurIPS 2019 Workshop on Safety and Robustness in Decision Making},
  year = {2019}
}