|
Abstract:
|
As an emerging machine learning paradigm, deep learning has quickly become a powerful tool in many areas such as image processing, natural language recognition, bioinformatics, and drug development. The performance of different deep learning methods, however, can be affected by various factors, such as architecture of a deep network, method of unsupervised pre-training, number of hidden layers, number of nodes in hidden layers, form of activation function, method of optimization, method of weight decay, dropout rate, and method of weight initialization. In many applications these parameters are often set in an ad hoc manner and thus it may not be clear whether an adopted deep learning method would render optimal performance. Therefore, there is a need for a systematic investigation of the impact of these parameters on the performance of deep learning methods. In this study, we will conduct a comprehensive simulation study in the context of multi-category classification to evaluate the performance of deep learning methods under various combinations of learning parameter configurations.
|
