Unlock instant, AI-driven research and patent intelligence for your innovation.

Isovariant network training method and device, electronic equipment and storage medium

A network training and network technology, applied in the field of artificial intelligence, can solve problems such as limited equivariance, achieve the effect of strengthening generalization, reducing learning space, and improving decoupling ability

Active Publication Date: 2022-05-24
北京智源人工智能研究院
View PDF8 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

It can be seen that because the existing equivariant networks are all discrete equivariant realized on discrete groups, their equivariance is very limited.

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Isovariant network training method and device, electronic equipment and storage medium
  • Isovariant network training method and device, electronic equipment and storage medium
  • Isovariant network training method and device, electronic equipment and storage medium

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0043] figure 1 A flowchart of an embodiment of an equivariant network training method according to an exemplary embodiment of the present invention, as shown in figure 1 As shown, the equivariant network training method includes the following steps:

[0044] Step 101: Construct an equivariant network composed of equivariant convolutional layers.

[0045] In this embodiment, the equivariant convolution layer is used to perform an equivariant convolution operation on the input feature map or image, and the equivariant condition that the equivariant convolution operation needs to meet is: [L g [f⊙Ψ]](x)=[[L g f]⊙Ψ]; where f(x) represents the input image or feature map, x is the spatial position (such as two-dimensional space or higher-dimensional space), L g is the transformation on the transformation group G, g∈G, ⊙ is the symbol of the equivariant convolution operation.

[0046] Optionally, the equivariant network structure may adopt G-CNN (Group-equivariant Convolutional ...

Embodiment 2

[0066] figure 2 for the invention figure 1 A schematic diagram of the construction process of an isovariable network shown in the embodiment shown, image 3 for the invention figure 1 A schematic diagram of the training process of an isovariable network shown in the embodiment shown, the following is combined with figure 2 and image 3 As shown, the training process of the equivariant network is introduced in detail:

[0067] First, as figure 2 As shown, build an equivariant network consisting of equivariant convolutional layers and replace the equivariant convolutional layers with Gaussian modulated equivariant convolutional layers.

[0068] In the embodiment of the present application, Gaussian modulation is performed on the equivariant convolution layer in the equivariant network. Since the traditional convolution layer is defined in discrete space, only the sampling grid points have parameters, and after Gaussian modulation The equivariant convolution layer is der...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses an equivariant network training method and device, electronic equipment and a storage medium. The method comprises the steps that an equivariant network composed of equivariant convolutional layers is constructed; performing Gaussian modulation on an equivariant convolutional layer in the equivariant network; and training the equivariant network by using the images in the training image set until the convergence of the equivariant network is finished. Since a traditional convolutional layer is defined in a discrete space and only sampling lattice points have parameters, Gaussian modulation is carried out on the equivariant convolutional layer in the equivariant network, the modulated equivariant convolutional layer is derivable in spatial position and angle, and position and angle parameters can be optimized, so that the parameters can be defined in a continuous space, and the spatial position and angle of the equivariant convolutional layer can be optimized. Therefore, the trained equivariant network can overcome the limitation of discrete sampling, so that the equivariant property on the continuous transformation group is obtained.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to an equivariant network training method, device, electronic device and storage medium. Background technique [0002] Usually, the rotation, scaling and other transformations of visual objects in the image will change the appearance of the visual objects, which will bring a lot of learning space to the neural network. The current feasible solution is to enhance the decoupling ability of the neural network to obtain a more compact potential learning space. The convolutional neural network has the inherent translational decoupling ability, and can learn basic features regardless of the position of the object in the input. In order to further improve the decoupling capability of the network, group equivariant theory and rotating group equivariant network have been proposed. [0003] In practice, whether it is an input image, a feature map, or a convolution kernel in t...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06N3/04G06N3/08
CPCG06N3/08G06N3/045
Inventor 陈智强余山陈阳
Owner 北京智源人工智能研究院