Method and device for model testing

Abstract

The embodiment of this specification provides a method and device for model testing. According to this method, first obtain samples from the test sample set, and then input the samples into multiple models to be tested included in the model set to obtain the output results of each model to be tested , and then determine the test result according to the output result, further, if the test result does not meet the predetermined condition, according to the predetermined rule, generate a new sample based on the above sample, and add the generated new sample to the test sample set. In this way, when the model testing method is executed cyclically, on the one hand, the accuracy and / or test adequacy of the model to be detected will be evaluated; samples to improve the effectiveness of model testing.

Description

technical field [0001] One or more embodiments of this specification relate to the field of computer technology, and in particular, to a method and device for testing a model by a computer. Background technique [0002] With the development of computer and artificial intelligence technology, there are more and more applications of artificial neural network (ANN), such as pattern recognition, automatic control, signal processing, auxiliary decision-making and so on. An artificial neural network is a computational model consisting of a large number of processing units, or neurons, connected to each other. Each processing unit represents a specific output function called an activation function. The performance test of the neural network model often passes the code coverage rate of the system that generates the neural network model, or its output accuracy rate on the sample set, and the output under special circumstances where the sample is difficult to cover (such as a tempera...

AI Technical Summary

Problems solved by technology

However, this method wastes manpower by manually collecting samples. At the same time, the collected training samples are limited after all, which cannot guarantee the full testing of the model, and the improvement of the accuracy of the model is also limited.

For example, the temperature in a certain place has never been as low as minus 70 degrees Celsius, and no amount of samples collected can cover the situation where the temperature is minus 70 degrees Celsius

In particular, in some special cases, such as an unmanned vehicle facing a truck with a white background, if there is no such situation in the collected test samples, it may be judged to be the sky background, affecting decision-making, which may cause irreparable losses

Images