Exploiting Structures in Weight Matrices for Efficient Real-Time Drone Control with Neural Networks

Matthias Kissel, Sven Gronauer, Mathias Korte, Luca Sacchetto, Klaus Diepold

Abstract: We consider the task of using a neural network for controlling a quadrotor drone to perform flight maneuvers. For that, the network must be evaluated with high frequency on the microcontroller of the drone. In order to maintain the evaluation frequency for larger networks, we search for structures in the weight matrices of the trained network. By exploiting structures in the weight matrices, the propagation of information through the network can be made more efficient. In this paper, we focus on four structure classes, namely low rank matrices, matrices of low displacement rank, sequentially semiseparable matrices and products of sparse matrices. We approximate the trained weight matrices with matrices from each structure class and analyze the flying capabilities of the approximated neural network controller. Our results show that there is structure in the weight matrices, which can be exploited to speed up the inference, while still being able to perform the flight maneuvers in the real world. The best results were obtained with products of sparse matrices, which could even outperform non-approximated networks with the same number of parameters in some cases.