Challenges and opportunities in the supervised learning of quantum circuit outputs

Abstract

Recently, deep neural networks have proven capable of predicting some output properties of relevant random quantum circuits, indicating a strategy to emulate quantum computers alternative to direct simulation methods such as, e.g., tensor-network methods. However, the reach of this alternative strategy is not yet clear. Here we investigate if and to what extent neural networks can learn to predict the output expectation values of circuits often employed in variational quantum algorithms, namely, circuits formed by layers of CNOT gates alternated with random single-qubit rotations. On the one hand, we find that the computational cost of supervised learning scales exponentially with the inter-layer variance of the random angles. This allows entering a regime where quantum computers can easily outperform classical neural networks. On the other hand, circuits featuring only inter-qubit angle variations are easily emulated. In fact, thanks to a suitable scalable design, neural networks accurately predict the output of larger and deeper circuits than those used for training, even reaching circuit sizes which turn out to be intractable for the most common simulation libraries, considering both state-vector and tensor-network algorithms. We provide a repository of testing data in this regime, to be used for future benchmarking of quantum devices and novel classical algorithms.