High-Throughput Screening And Genetic Search For Audio Feature Discovery

Appropriate features are critical to the performance of machine learning algorithms, but feature engineering can require years of research on a specific problem domain before a solution is found that performs well. One approach to automating this task is to use deep learning algorithms for feature discovery. A problematic aspect of this approach is that the performance of deep learning algorithms depends on the settings of a number of model parameters that are not well understood.
In this thesis, I discuss the use of high-throughput screening to find effective parameter instantiations for deep learning methods. I apply this methodology to the audio domain, on which it produces models that perform with 61% accuracy on a language classification task, compared to a baseline of 52%. I improve upon the random search used in high-throughput screening literature by incorporating a genetic search that produces models that perform with 66% accuracy on this task.