Method and device for enhancing fault tolerance and fluctuation robustness in extreme situations

Abstract

The invention provides a method for checking the integrity of parameters of a neural network and saving computing resources. The method comprises the following steps that when information about each of a (t-2)-th image and a (t-1)-th image corresponding to a (t-i2)-th frame and a (t-i1)-th frame of a video is acquired, a calculation device causes a background prediction unit to generate t-th background prediction information about a t-th image by referring to the information about each of the (t-2)-th image and the (t-1)-th image; when the t-th image is acquired, the calculation device causesthe pattern insertion unit to generate at least one inspection input by referring to the t-th background prediction information and inserting the test pattern into the t-th image. A calculation devicecauses the neural network to generate at least one inspection output by applying a convolution operation to the inspection input; and the calculation device causes the inspection unit to determine the integrity of the neural network by referring to the inspection output and at least one reference output corresponding to the test pattern.

Description

technical field [0001] The present invention relates to a method and a device capable of saving computing resources consumed in the process of using test patterns to verify the integrity of convolution parameters. In more detail, it relates to a method of checking the integrity of more than one parameter of a neural network and saving more than one computing resource by inserting at least one test pattern into at least one background region of at least one input image, comprising the following steps: (a) After acquiring information about each of the (t-2)th image and the (t-1)th image respectively corresponding to the (t-i2)th frame and the (t-i1)th frame of the image In this case, the computing device causes at least one background prediction unit to generate the t-th background prediction information of the t-th image by referring to information about each of the (t-2)-th image and the (t-1)-th image (b) in the case of acquiring the tth image, the computing device causes at...

AI Technical Summary

Problems solved by technology

[0005] Therefore, a method is needed to check whether the CNN parameters maintain the same parameters as when the CNN completed the initial learning during the test process, but the reality is that the research so far has mainly focused on how well the CNN for autonomous driving can move the vehicle, and how to There is little research on methods of maintaining this safety

[0006] In addition, even if a surveillance system that maintains the safety of CNN is created using existing technologies in other fields, there is a disadvantage that the performance of the original function of CNN may be degraded due to many additional calculation processes other than the CNN calculation process.

Images