Abstract:
In this paper, we present a new binarization approach of point-wise operations based on Expectation-Maximization (POEM) for efficient deep learning on point clouds. In this work, we first implement a powerful baseline binarization method and discover that two main drawbacks cause the immense performance drop layer-wise Gaussian-distributed weights and non-learnable scale factor. Based on the analysis, we formulate our POEM to solve problems in two steps. First, we propose a novel weight optimizer based on the Expectation-Maximization (EM) algorithm to constrain weights to formulate a robust bi-modal distribution. Moreover, a well-designed reconstruction loss is introduced to calculate learnable scale factors to enhance the representation capacity of 1-bit fully-connected (Bi-FC) layers. Extensive experiments demonstrate that our POEM surpasses existing binarization methods by significant margins.
