A Quantum-Inspired Algorithm
After “Quantum MNIST” with dense angle encoding and “784-Dimensional Quantum MNIST” with amplitude encoding, I experimented with “Quantum-Inspired MNIST” using a classical interpretation of the quantum algorithm that can be used for Quantum Machine Learning (QML) tasks such as quantum classification and quantum clustering. Using nothing but addition and subtraction, the approach achieved 72% accuracy, which is much better than guessing, better than some Convolutional Neural Network (CNN) student attempts, and yet noticeably worse than Kaggle submissions.
This is my first attempt to boost the accuracy of Quantum-Inspired MNIST, while hopefully retaining the relative simplicity of Quantum-Inspired MNIST.
MNIST Dataset
This begs the question, “what’s MNIST?” For those unfamiliar with it, MNIST is a popular dataset of handwritten digits, the numbers 0 through 9. It is popular because it is thoroughly tested and clean, which allows students to focus on building and training their image classification models instead of on their data.
Each digit in the MNIST dataset is represented by approximately 6,000 images. Each image contains 784 pixels arranged in a 28x28 grid. Each pixel has a value ranging from 0 to 254, representing its intensity. I have…
