Model-universal deep neural network representation visualization method and device

Abstract

In order to solve the problem of poor releasability of a convolutional neural network in the prior art, the invention provides a convolutional neural network representation visualization method, a convolutional neural network representation visualization device and a training method, so as to improve the releasability of the convolutional neural network. The convolutional neural network representation visualization method comprises the steps: obtaining a first feature map after a to-be-visualized image is input into a convolutional neural network; depooling the first feature map to obtain a second feature map; correcting the second feature map through a ReLU function to obtain a third feature map; performing deconvolution or guided back propagation on the third feature map to obtain a first visual feature map; and displaying the first visual feature map. The invention further discloses a corresponding visualization device and a convolutional neural network training method based on therepresentation visualization method. Through depooling, deactivation, deconvolution and guided back propagation operation, the feature activation condition of the to-be-visualized layer in the convolutional neural network model is displayed, so the releasability of the convolutional neural network is improved.

Description

technical field [0001] The present disclosure relates to the field of intelligence, and in particular to a convolutional neural network representation visualization method, device and training method. Background technique [0002] Although the existing convolutional neural network can be an "end-to-end" model trained by constructing and labeling large data for error backpropagation to optimize parameters; it can have gratifying performance in some scenarios, but these convolutional neural network models are There is a similar problem that cannot be ignored: poor interpretability. That is, although the model gives a high accuracy rate, it cannot give more reliable information to explain the consideration basis for the given results. Such a situation has resulted in industries that require high interpretability, such as finance, medical care, and autonomous driving, in specific fields. Although the result data of the convolutional neural network model looks good, it cannot be...

AI Technical Summary

Problems solved by technology

[0002] Although the existing convolutional neural network can be an "end-to-end" model trained by constructing and labeling large data for error backpropagation to optimize parameters; it can have gratifying performance in some scenarios, but these convolutional neural network models are has the same problem that cannot be ignored: poor interpretability

That is, although the model gives a high accuracy rate, it cannot give more reliable information to explain the consideration basis for the given results

This situation has caused that in specific fields such as finance, medical care, and autonomous driving, industries that require high interpretability, although the result data of the convolutional neural network model looks good, it cannot be practically implemented in applications.

At the same time, poor interpretability means that people cannot trust the answers given by the model when faced with new data, which leads to a great reduction in the predictive function of the model, and further limits the practical application of deep models in various industries

Images