Our team has recently demonstrated that a quantum reservoir neural network implemented on a circuit QED system composed of a transmon qubit coupled to a superconducting cavity can learn to classify input classical data. In this pilot experiment, neural outputs were obtained by measuring probability occupations of different Fock states of a single quantum oscillator and training was performed on a classical computer after the measurement. In order to perform harder learning tasks and increase the expressivity of the neural network, training should be done in the quantum system as well. For this, we are developing new training algorithms, specific to analog quantum systems. The goal of the internship and subsequent PhD thesis is to simulate and implement layers of parametrized operations that will be applied on the quantum systems and whose parameters will be trained using physics aware learning methods.