Quantum Machine Learning using a Dissipative Quantum Reservoir

Recurrent neural networks (RNNs) are often slow learners, requiring extensive training of their hidden layers. In reservoir computing, the RNN's hidden layers are replaced by a reservoir, which can be a complex circuit (e.g., echo state network, or ESN) or physical system. The reservoir transforms recent temporal data into output patterns that can be read by a trainable single layer of neurons. The quantum reservoir, with just a few qubits and including dissipation, has recently been shown outperform a much larger classical ESN. Here we discuss the charge density wave (CDW) - a correlated electron-phonon system - as a candidate quantum reservoir. Some CDW materials (e.g., NbSe3) show learning, such as a pulse-duration memory effect, where 1-3 training pulses are needed experimentally vs. 100's to 1000's in classical simulations. This occurs in a highly dissipative environment with many normal electrons. Related materials (e.g., NbS3) have optimum CDW transport properties at room temperature, suggesting the possibility of certain types of quantum machine learning at high temperatures.