Quantum Classifier with Iterative Re-Uploading for Universal Classification: Performance Evaluation and Insights

In this work, we leverage the capabilities of an existing quantum classifier known for its ability to perform multi-class classification of multi-dimensional data using a few qubits. Our approach involves the iterative re-uploading of classical data into the quantum circuit, punctuated with parametrized rotation gates acting as processing units. We evaluated our implementation across various classification tasks of differing complexity levels. In our experiments, we employed several minimization methods, including SLSQP, CG, and Nelder-Mead, to observe variations in accuracy for each problem. We also experimented with random datasets in addition to fixed ones and introduced an additional cost function, trace distance. Furthermore, we added a ‘line’ problem. For the linear problem, we achieved high accuracy (around 0.92) for the fidelity cost function using both random and fixed datasets with SLSQP and CG methods. For the non-linear classification problem, we observed the highest accuracy for fidelity and fixed dataset (0.97). We also achieved good accuracy (up to 0.91) for trace distance for linear problems using the SLSQP method and random data set. Remarkably, all these accuracies were achieved with only 50 training data points.