There have been many attempts at translating classical machine learning algorithms to quantum circuits. Our architecture utilizes generalized 3-qubit parametrized unitary gates to perform image classification. However, simply copying the architecture to use quantum circuits does not result in similar transformations, so we introduce an encoding scheme that allows us to preserve the quasilocal transformations of a convolutional neural network. The results are promising, with 80% accuracy on the 10-class MNIST dataset, with further improvements from decreasing the number of classes, leading up to a 99% accuracy on a 2-class dataset. However, it is important to keep in mind that image classification is not necessarily a task that benefits from a quantum algorithm due to encoding constraints. MNIST classification is simply a benchmark to show the feasibility of our architecture and compare it to existing algorithms, quantum and classical. There is further ongoing research to apply this architecture to problems that are more suited to quantum computing.
Find the code for this project here.