A method for constructing a high-dimensional convolution acceleration neural network

Abstract

The invention provides a method for constructing a high-dimensional convolution acceleration neural network. The method comprises the following steps: explaining a currently popular Gaussian convolution acceleration operation SBS as a convolution layer operation of the neural network; carrying out extended tensor decomposition (CP decomposition) to gCP decomposition to provide corresponding g-convolution; The gCP layer is further expanded to the cascaded gCP layer as gHT decomposition, and the expressive force of the network is improved; Selecting an activation function g, calculating gradients of the g and the sum function, and training the AccNet through a back propagation algorithm; and extracting spring operation, spring operation and spring operation on the AccNet of the training element to form an optimal acceleration algorithm of any high-dimensional convolution.

Description

technical field [0001] The invention relates to a high-dimensional convolution technology, in particular to a method for constructing a high-dimensional convolution accelerated neural network. Background technique [0002] High-dimensional convolution is widely used in various disciplines. For example, cascaded convolution is used to represent generalized Laplacian distance in visual matching, high-dimensional Gaussian convolution is used in fully connected CRFs inference for effective information transfer, and bilateral filtering transforms into convolutions in higher-dimensional spaces of increasing dimensionality, and so on. However, due to the high computational complexity of high-dimensional convolution, there are serious performance problems, so in the past few decades, people have used manual methods to design fast algorithms for Gaussian convolution, but recently for various non-Gaussian convolution The need for convolution has emerged, and the controversy has becom...

AI Technical Summary

Problems solved by technology

Due to the separable blur kernel, the d-D Gaussian blur performed on these vertices can be regarded as the sum of several separable 1-D filters, so the computational complexity of a d-D convolution (1) varies from O(r d ) is reduced to O(rd), and because the filtering window becomes very small after splatting, the computational complexity can be roughly regarded as O(d) independent of the radius r, but this algorithm cannot approach non-Gaussian convolution and how to improve Approximation error has no theoretical basis

Images