A physics inspired class of quantum neural network for generative training
TITLE: Training Fully Quantum Boltzmann Machines
SPEAKER: Affiliate Assistant Professor Nathan Wiebe
AFFILIATION: Pacific Northwest National Laboratory | University of Washington, USA
HOSTED BY: Dr Márika Kieferová, UTS Centre for Quantum Software and Information
ABSTRACT:
In recent years quantum machine learning has grown by leaps and bounds but a major problem still vexes the field is how to efficiently train quantum neural networks. This is particularly challenging because of the lack of a natural backpropagation algorithm for updating the quantum model.
In this talk, I will focus on an approach that can mitigate this problem through generative training. We will show how to construct a fully quantum model of a Boltzmann machine and train all of the parameters of that model for both the quantum and classical parameters in the model. In contrast, existing methods were not able to achieve this.
In particular, we will show explicit query upper bounds for the cost of simulation, provide a formal proof for BQP-completeness for evaluating such neural networks and also discuss remaining problems in the field and how to generalize the ideas presented here to go beyond Boltzmann machines to allow efficient training of broad classes of quantum neural networks.
REFERENCES:
Tomography and Generative Data Modeling via Quantum Boltzmann Training: https://arxiv.org/abs/1612.05204
Generative training of quantum Boltzmann machines with hidden units: https://arxiv.org/abs/1905.09902
Pacific Northwest National Labs: https://www.pnnl.gov/
UTS Centre for Quantum Software and Information: https://www.uts.edu.au/research-and-teaching/our-research/centre-quantum-software-and-information
Márika Kieferová: https://www.uts.edu.au/staff/maria.kieferova
