Abstract (300 words maximum)
A common practice for developing a Neural Network architecture is to build models in which each layer has a single activation function that is applied to all nodes uniformly. This paper explores the effects of using multiple different activation functions per layer of a neural network to analyze the effects of the architectures’ ability to generalize datasets. This approach could allow neural networks to better generalize complex data providing better performances than networks with uniform activation functions. The approach was tested on a fully connected neural network and compared to traditional models, an identical network with uniform activations, and an identical network with a different activation function for each layer. The models are tested on several problem types and datasets and analyzed to compare the performance and training time.
Academic department under which the project should be listed
CCSE - Computer Science
Primary Investigator (PI) Name
Md Abdullah Al Hafiz Khan
Analysis of Multi-Activation Layers in Neural Network Architectures
A common practice for developing a Neural Network architecture is to build models in which each layer has a single activation function that is applied to all nodes uniformly. This paper explores the effects of using multiple different activation functions per layer of a neural network to analyze the effects of the architectures’ ability to generalize datasets. This approach could allow neural networks to better generalize complex data providing better performances than networks with uniform activation functions. The approach was tested on a fully connected neural network and compared to traditional models, an identical network with uniform activations, and an identical network with a different activation function for each layer. The models are tested on several problem types and datasets and analyzed to compare the performance and training time.