Design of a deep neural network suitable for real-time feedback strategies discovered via reinforcement learning on a quantum device

Real-time feedback for quantum systems is an essential ingredient in many quantum control tasks, such as quantum error correction and quantum state preparation. Two aspects make feedback challenging: First, a time far shorter than the coherence time is required. Second, it represents a complex decision making problem. The subfield of machine learning dealing with optimizing strategies for problems of this type is reinforcement learning, whose power has been convincingly demonstrated in areas ranging from robotics to video and board games.

In this work, we present and analyze the design of a neural network which can be implemented on a Field Programmable Gate Array (FPGA) to discover real-time feedback strategies for initialization of a superconducting qubit. Our network will be trained via reinforcement learning which is only based on data directly accessible in an experiment, such that precise knowledge of the underlying dynamics is not required. To address the challenge of low latency, we introduce a key idea: the network inference computation is interleaved with the simultaneous collection of measurement data.