Tuning the Parameters of a Convolutional Artificial Neural Network by Using Covering Arrays

Artificial Neural Networks have proven to be a very powerful machine learning algorithm which can be adequate to learn successfully a variety of tasks. Currently, very complex classification problems on different kind of data (images, video, sound, text, DNA) have been solved using neural networks. This kind of algorithms usually has many parameters that need to be fine-tuned in order to have good results. Usually, this tuning is made by trial and error. However, this procedure does not guarantee the optimal performance of the training process. In this work, we study the use of mixed-level covering arrays to design experiments that help us to find the best combinations of parameters. We tested this approach by tuning a convolutional neural network for an audio classification task. For the implementation, we took advantage of the flexibility of the open source software library for machine learning TensorFlow.