Improving deep neural networks using softplus units

Recently, DNNs have achieved great improvement for acoustic modeling in speech recognition tasks. However, it is difficult to train the models well when the depth grows. One main reason is that when training DNNs with traditional sigmoid units, the derivatives damp sharply while back-propagating between layers, which restrict the depth of model especially with insufficient training data. To deal with this problem, some unbounded activation functions have been proposed to preserve sufficient gradients, including ReLU and softplus. Compared with ReLU, the smoothing and nonzero properties of the in gradient makes softplus-based DNNs perform better in both stabilization and performance. However, softplus-based DNNs have been rarely exploited for the phoneme recognition task. In this paper, we explore the use of softplus units for DNNs in acoustic modeling for context-independent phoneme recognition tasks. The revised RBM pre-training and dropout strategy are also applied to improve the performance of softplus units. Experiments show that, the DNNs with softplus units get significantly performance improvement and uses less epochs to get convergence compared to the DNNs trained with standard sigmoid units and ReLUs.