Quantum advantage in training binary neural networks

Neural networks have become a corner stone of artificial intelligence. Once a task is defined, a neural network needs to undergo a training phase in order to calibrate the network parameters. On a classical computer, the number of training cycles is strongly correlated with the amount of parameters in the neural network, making an optimization of the parameters computationally expensive. Here, we propose the a fully quantum training protocol for quantum binary neurons. We show that for special cases this protocol yields a quadratic advantage over commonly used classical training methods, and numerics suggests that this advantage is generic for most instances. The source of this advantage is the possibility to run training cycles in a quantum superposition. Further, we present an extension of the training of single quantum binary neurons to feedforward binary neural networks.

PDF Abstract