BUP9 - RoNNi

Improving the robustness and energy efficiency of neural networks through equivalence testing.

Machine learning is used today in both manufacturing and vehicle control. Neural networks play a central role in the field of autonomous driving. Their use in this context is hampered by high memory and energy requirements as well as by a lack of safety guarantees.

Equivalence checking allows to visualize (undesired) deviations in updates of neural networks and to generate targeted training data to minimize the deviation between two networks. In particular, formal methods can be used to verify whether downsized and thus more energy-efficient networks still meet the original quality requirements.

Approach

Definition of first attenuated equivalence terms for neural networks
Description of the relevant input space of neural networks, such as valid parameters of a motor control system
Development of a method for equivalence checking based on Constraint Programming and Abstract Interpretation
Implementation of a software prototype for the developed method
Optimization of the method on relevant neural networks