Motivation Deep neural network (DNN) algorithms were utilized in predicting various biomedical phenotypes recently, and demonstrated very good prediction performances without selecting features. This study proposed a hypothesis that the DNN models may be further improved by feature selection algorithms. Results A comprehensive comparative study was carried out by evaluating 11 feature selection algorithms on three conventional DNN algorithms, i.e. convolution neural network (CNN), deep belief network (DBN) and recurrent neural network (RNN), and three recent DNNs, i.e. MobilenetV2, ShufflenetV2 and Squeezenet. Five binary classification methylomic datasets were chosen to calculate the prediction performances of CNN/DBN/RNN models using feature selected by the 11 feature selection algorithms. Seventeen binary classification transcriptome and two multi-class transcriptome datasets were also utilized to evaluate how the hypothesis may generalize to different data types. The experimental data supported our hypothesis that feature selection algorithms may improve DNN models, and the DBN models using features selected by SVM-RFE usually achieved the best prediction accuracies on the five methylomic datasets. Availability and implementation All the algorithms were implemented and tested under the programming environment Python version 3.6.6. Supplementary information Supplementary data are available at Bioinformatics online.

Weight tying is now a common setting in many language generation tasks such as language modeling and machine translation. However, a recent study reveals that there is a potential flaw in weight tying. They find that the learned word embeddings are likely to degenerate and lie in a narrow cone when training a language model. They call it the representation degeneration problem and propose a cosine regularization to solve it. Nevertheless, we prove that the cosine regularization is insufficient to solve the problem, as the degeneration is still likely to happen under certain conditions. In this paper, we revisit the representation degeneration problem and theoretically analyze the limitations of the previously proposed solution. Afterward, we propose an alternative regularization method called Laplacian regularization to tackle the problem. Experiments on language modeling demonstrate the effectiveness of the proposed Laplacian regularization.

No abstract

No abstract

No abstract