Quantum Model Recycling: Turning a Classifier into a Generator

In conventional machine learning practice, generative and predictive learning are treated as independent problems, and models for each are trained separately with distinct architectures and methodologies. Here, we show that the bit-bit encoding architecture in quantum machine learning gives rise to learned classifiers $U$ which can be used to build Quantum Circuit Born Machines (QCBM) which act as conditional generators. The QCBMs utilize $U$ as a generalized oracle in a Grover-type iteration. Constructing generators in this way does not require any additional training. We demonstrate on a subset of the MNIST dataset.