Information geometry based extreme low-bit neural network for point cloud

Abstract

Deep learning has significantly advanced three-dimensional computer vision applied to point clouds. Nevertheless, the substantial consumption of time, storage, and energy substantially limits its deployment on edge devices with constrained resources. Extremely low bit quantization has received wide attention due to its extremely high compression ratio, but the problem of a significant drop in accuracy cannot be ignored. To alleviate the obvious accuracy degradation for extreme low-bit quantization, this paper proposes a novel compression framework for binary and ternary neural networks applied to point clouds, which introduces information geometry to model quantization to compensate the severe feature manifold distortion. It applies differential geometry on manifolds to study the implicit information of the point cloud feature data. Based on the theoretical analysis from the novel perspective of information geometry, two optimization modules are proposed to alleviate severe geometry distortions on differential manifolds. The first module, scaling recovery, provides layer-wise scaling parameters to reduce geometric distortion caused by quantization. The second module, Pooling Recovery, is specially designed to alleviate more severe pooling geometry distortions in point clouds. These two modules benefit both binary and ternary neural networks with ignored overheads. For ternary quantization, optimizations on convolution weights and gradients are additionally introduced. The proposed self-adaptive gradient estimation provides a more accurate approximation to the non-differential ternary staircase function. Convolution weight optimization is implemented on an information-geometry optimized model to achieve even higher accuracy and less memory consumption. Experimental results validate that the proposed models significantly outperform state-of-the-art methods and demonstrate better scalability. Overall, this compression framework has the potential to facilitate the deployment of deep learning models on edge devices with limited resources, opening up new opportunities for applications of three-dimensional computer vision.