EvoDNN - Evolving Weights, Biases, and Activation Functions in a Deep Neural Network

Classification, such as classifying cell samples into cancer (malignant) or normal (benign), or the classification of genome regions into functional regions (for example coding regions or promoter regions) are important problems in Computational Biology. For such tasks, we have previously designed an evolutionary deep neural network that in addition to evolving the neuron's weights and biases also evolves (learns) the activation functions for each neuron and called this approach EvoDNN. EvoDNN can employ activation functions that are non-differentiable as it does not rely on back-propagation. This feature is adding flexibility in terms of activation functions EvoDNN can employ. The work presented here extends our previous work on EvoDNN by analyzing a more extensive set of data sets and studying variations of the internal topology of the EvoDNN model. In addition, we study the effect of evolving the weights and biases only while holding the activation function fixed and demonstrate that evolving activation functions indeed provides better performance. We also compare our model to several other popular models and demonstrate superior performance on several data sets. The current code for EvoDNN is available at https://github.com/Payuing/evoDNN.