LAUSR

Winograd's minimal filtering algorithm effectively reduces the multiplication arithmetic complexity of Convolutional Neural Networks. However, Winograd convolutions in current implementations are limited to small feature tiles for two reasons: the numerical error and the overhead of the transformations. The performance of Winograd convolutions is determined by the points used to construct transformation matrices, which raises a great challenge to find the optimal points: it requires exploring the vast design space trading off between numerical accuracy and hardware resource consumption. In this letter, we introduce the Hardware-Aware point selection framework for efficient Winograd convolution, which leverages reinforcement learning to determine the point selection policy. We design three reward functions to optimize numerical accuracy and circuit area. Experiments demonstrate that Winograd convolutions using our policies outperform state-of-the-art methods in circuit area and accuracy.