Component Extraction for Deep Learning Through Progressive Method.

Machine learning has shown great impact in a lot of applications. Within all types of tools, deep learning should be one of the most important techniques thanks to its ability to capture the correlation between the input features and output results. However, the relatively long training time and high computation complexity remain a big problem in deep learning. In addition, the impossibility to explain the model makes it harder for us to look for alternatives to fix the bad fitting results. Therefore, this paper aims at improving the deep learning model training result by proposing a principal component extraction algorithm. Compared with the previous Principal Component Analysis (PCA) methods, this algorithm creatively consider not only the original input components but also the computed variables in the first hidden layer in neural network so as to capture more representative components. The experiment shows that compared to previous PCA method, this can better capture the principal components from all input variables.