Solving MNIST with a globally trained Mixture of Quantum Experts

TL;DR

This paper introduces a quantum neural network architecture that classifies MNIST images with high accuracy using amplitude encoding and a mixture of experts, demonstrating the potential for scalable quantum machine learning.

Contribution

The paper presents a novel quantum neural network architecture based on a mixture of experts that efficiently encodes full-resolution images and improves trainability for high-accuracy classification.

Findings

Achieved 97.5% accuracy on MNIST parity classification

Efficient amplitude encoding of full-resolution images with 10 qubits

Training all experts jointly enhances model trainability

Abstract

We propose a new quantum neural network for image classification, which is able to classify the parity of the MNIST dataset with full resolution with a test accuracy of up to 97.5% without any classical pre-processing or post-processing. Our architecture is based on a mixture of experts whose model function is the sum of the model functions of each expert. We encode the input with amplitude encoding, which allows us to encode full-resolution MNIST images with 10 qubits and to implement a convolution on the whole image with just a single one-qubit gate. Our training algorithm is based on training all the experts together, which significantly improves trainability with respect to training each expert independently. In fact, in the limit of infinitely many experts, our training algorithm can perfectly fit the training data. Our results demonstrate the potential of our quantum neural network to achieve high-accuracy image classification with minimal quantum resources, paving the way for more scalable and efficient quantum machine learning models.