Seminar: Yidong Liao, UTS QSI

Title: Quantum advantage in training binary neural networks   

Presenter: Yidong Liao from UTS, Centre for Quantum Software and Information

Abstract: Training classical neural networks is computationally costly, and local optimisation methods like gradient descent may give a local minimum of the cost function. This talk will present an explicit algorithm for using quantum search techniques to find global optimum in a time that is quadratically -in the number of weight configurations- faster than a brute force search. To achieve this we define quantum versions of binary neural networks. Binary Neural Networks (BNNs) are neural networks with weights and activations limited to take only binary values ±1. BNNs have been shown to achieve state-of-the-art performance in memory saving and inference time, which enables the usage of deep neural networks to expand from running in the cloud remotely to performing local inference on resource-constrained devices (mobile phones, IoT devices, robots etc.). In our training algorithm for BNN, the training acts as quantum search on a superposition of weight states, which guaranteed to find global optimum for any cost function landscape. A number of examples of the circuit implementation for the algorithm will be presented.