Fast-QTrain: an algorithm for fast training of variational classifiers

This work proposes a novel algorithm (Fast-QTrain) that enables fast training of variational classifiers. This training speedup is achieved by processing multiple samples, from a classical data set, in parallel during the training process. The proposed algorithm utilizes a quantum RAM along with other quantum circuits for implementing the forward pass. Besides, instead of computing the loss classically, which is the usual practice, we calculate the loss here using a swap test circuit. As a result, our algorithm reduces the training cost of a variational classifier trained for m epochs from the usual 𝒪(mN) (which is also the case with most classical machine learning algorithms) to 𝒪(N + mlog N) where the data set contains N samples. Ignoring the one-time overhead of loading the N training samples into the qRAM, the time complexity per epoch of training is 𝒪(log N) in our proposed algorithm, as opposed to 𝒪(N) (which is the case for other variational algorithms and most classical machine learning algorithms). By performing quantum-circuit simulations on the Pennylane package, we show fairly accurate training using our proposed algorithm on a popular, classical data set: Fisher’s Iris data set of flowers. While we restrict ourselves to binary classification (of samples from classical data sets) in this paper, our algorithm can be easily generalized to carry out multi-class classification. Our proposed algorithm (Fast-QTrain) can also be adapted for any classification ansatz used in the variational circuit as long as the encoding of the classical data into qubits is non-parameterized.