Depth Learning Standard Deviation Loss Function

作者：Chunlin Wang, Jianyong Sun, Wanjin Xu, Xiaolin Chen · 发表于：Journal of Physics Conference Series · 年份：2019 · DOI：10.1088/1742-6596/1176/3/032050 · 被引用次数：5 · 研究领域：Advanced Neural Network Applications、Anomaly Detection Techniques and Applications、Domain Adaptation and Few-Shot Learning

Deep learning is a branch of the field of machine learning that outp-erforms humans in image and natural language processing. This mainly benefits from powerful computer processing capability and massive datasets. In the training model, the loss function needs to be used to evaluate each training result The quality of the loss function directly affects the correctness and validity of the model. In this paper, four kinds of common loss functions in deep learning are studied and our own loss function is proposed; Then the MNIST dataset is adopted to classify and train them; Finally, in the course of training, the change and the classification correct rate of the loss function value and model parameters are observed. Firstly, the different loss functions are experimented and their performance and application fields are analyzed, and secondly, our loss function is defined. After 30,000 times of iterative training, the test set data is used to test the loss function proposed, the correct rate reaches 98.53%, and achieves better accuracy in fewer training iterations, saving training times and resources.