Towards a quantum implementation of neural network verification algorithms

Description

Deep neural networks are the driving force behind the recent breakthroughs in AI. They are already widely deployed in everyday applications, including safety-critical domains (e.g., autonomous vehicles, medical diagnostics, and financial systems). The ability to formally verify the behaviour of these networks would provide substantial benefits, including reliability, safety and interpretability. Although many different algorithms for neural network verification are available, they suffer significant scaling challenges. As a result users have to compromise - either by reducing the size of the network which reduces its performance, or by simplifying the specification which less accurately captures the desired behaviour.

However, some of the most successful verification algorithms take the form of a highly structured branching search. This suggests that the problem is potentially well-suited to quantum computing, which is particularly effective at exploring exponentially many branches in parallel. This project would be to look at the initial stages of developing algorithms for neural network verification that are capable of running on quantum hardware.

Assumed Skills

Should have completed or about to start Quantum Computing and Machine Learning units.

Sample references

Paper describing the first neural network verification algorithm - https://arxiv.org/abs/1702.01135 Blog post explaining the algorithm in more detail - https://towardsdatascience.com/implementation-details-of-reluplex-an-efficient-smt-solver-for-verifying-deep-neural-networks-379ea359c41a