Study on Performance Improvement Techniques and Applications of Deep Learning

Abstract

Deep learning currently shows high performance in many real-life applications and has been applied to various environments, and research has been conducted. However, since deep learning is a black-box model, it is difficult to interpret it, and it is difficult to understand why it is getting better. Therefore, we will look at the research contents on the fields of application of deep learning performance improvement technology and deep learning to improve the performance of existing deep learning. To improve the performance of deep learning, we proposed the contents of the improvement of deep learning technology, which is improved by grasping where the problem is. Improved performance of deep learning technology, we look at seven types of technology. First, the bi-activation function: this is an enhancement activation function that is enhanced in the convolution neural network. Second, it is the loss of a neural network that is continuously occurring, and the continuous correction cascade loss occurs through continuous online learning. Third, non-linear regularization: This is an improved regularization version for the regularization method. Fourth, it is a new auxiliary component for optimizing the deep learning model. The fifth is ensemble normalization for stable learning. The sixth is the similarity analysis of actual fake fingerprints and fake fingerprints generated by DCGAN. Seventh is a multi-path decoder scheme with error reduction embedding in one-hot bi-directional Seq2Seq with adaptive regularization for music composition. In addition, deep learning that expresses high performance will be introduced to technology applied to real life. In technology using deep learning, we will look at four types of technology. The first is the importance of adaptive seeding. The second is the module comparison study in the image captioning. Third, the visualization of outlier data. Fourth, it is a stable and fine-grained segmentation that uses batch normalization and focal loss and L1 regularization in the U-Net structure. Through this, we will create a new deep learning theory using deep learning to improve the performance of the deep learning model and proceed to future research in the new research field.