Research On Neural Networks Producing High-confidence Predictions For Unknown Class Samples

With the explosive development of artificial intelligence in recent years,neural networks have obtained unprecedented performance in areas such as computer vision and have become the state-of-the-art standard for many learning tasks.Despite their great success,neural networks currently have some inherent limitations.For example,after training a neural network with normal supervision methods,neural network models usually perform well when the training and test data are independently sampled from the same distribution.However,when the distribution of the test data is different from the distribution of the training data,a neural network model may incorrectly classify an unknown class of samples not in the training set as a known class with high confidence,which may result in an extremely high error cost.Currently,it is difficult to collect training samples covering all the classes when training neural networks due to various limitations of uncontrollable reasons,so unknown class samples may be submitted to the model during testing,which requires that the model should not make high confidence predictions for samples far from the training data,which is especially important in safety-critical applications.In recent years,researchers have done a lot of research on this problem.One of the major approaches is to set up a binary detector to detect whether a sample is an unknown category,but this approach does not fundamentally solve the problem of neural networks generating high confidence in such unknown samples.Another method is to construct a subset of out-of-distribution samples from extra datasets and suppress the confidence of these out-of-distribution samples to achieve the purpose of suppressing the confidence of the prediction of other out-of-distribution samples.Combining the above problems,the main research work of this thesis is as follows.（1）In this thesis,a novel generative adversarial network is designed and trained to generate a subset of high-quality out-of-distribution unknown class samples without using extra data.（2）By suppressing the confidence of the out-of-distribution unknown class samples generated in this thesis during training,the neural network is able to make the network significantly alleviate the problem of high confidence of errors on many unknown samples and does not affect the performance of the original task.（3）The combination of the neural network trained by suppressing the confidence of the out-of-distribution unknown category samples generated in this thesis and the current state-of-the-art unknown class sample detection methods can further improve the detection performance.（4）In this thesis,we design a novel model training architecture that can effectively improve the accuracy of the original classification task by combining the generated out-of-distribution unknown classes samples.