Machine und Deep Learning Seminar

Permutations in Neural Networks and Quantum Annealing

Permutations arise in all applications with data that is not naturally ordered, for example the vertices in a point cloud which are stored in arbitrary order. Extracting the mapping between two instances is often a requirement for downstream tasks. However, optimizing for permutations is complicated due to their discrete nature and complicated constraints which do not scale well.  

In this talk, a new representation for permutations will be presented that enables continuous optimization that can be used in neural networks and furthermore does not require storage of a square permutation matrix. In addition, it will give an introduction to quantum G-annealing, which is extremely efficient in solving quadratic unconstrained binary optimization problems, and explain the possibilities for incorporating permutation constraints in this setting.