A method for testing the stability of deep neural networks

Abstract

The invention relates to a method for testing the stability of a deep neural network, comprising the following steps: inputting an initial test image into a deep neural network; performing multiple jitters on an input parameter of a certain layer to obtain a plurality of different input parameters, and then using the The calculation of this layer is performed with multiple parameters; the above steps are performed only in a certain layer or in multiple layers respectively, and continue to complete the calculation of the entire network to obtain multiple output results; multiple output results are accumulated and averaged as The final result; subtract the final result from the output before the jitter to calculate the difference, the smaller the difference, the higher the stability. The scheme of the present invention jitters the intermediate parameters of the deep neural network, not only can observe the stability of the neural network when dealing with some low-level and simple feature changes, but also can observe the stability of the deep neural network when dealing with high-level feature changes, Thus testing the neural network more comprehensively.

Description

technical field [0001] The invention belongs to the technical field of machine learning, and in particular relates to a method for testing the stability of a deep neural network. Background technique [0002] Deep neural networks have been widely used in the field of image processing. However, when a network is trained, the network performance is sometimes not stable enough in real use scenarios, because real-world images and training often have slight changes, which may be due to different models of image acquisition machines (optical cameras, CT, etc.), Or it may be the habit of the shooting operator, or it may be the influence of the environment (light conditions, etc.). However, it is often difficult to obtain comprehensive data changes in the real world. For example, in medical imaging, different hospitals often use different instruments, and each medical data acquisition has strict privacy issues. Therefore, when the model training is completed in the development pha...

AI Technical Summary

Problems solved by technology

First of all, this method needs to preprocess the image, which obviously increases a lot of calculation

Second, this method can only observe the stability of the model to some low-level feature changes, but cannot observe the stability of the model to high-level feature changes

Images