Online Deep Learning Hyperparameter Tuning based on Provenance Analysis

Training Deep Learning (DL) models require adjusting a series of hyperparameters. Although there are several tools to automatically choose the best hyperparameter configuration, user is still main actor take final decision. To decide whether training should continue or try different configurations, needs analyze online hyperparameters most adequate dataset, observing metrics such as accuracy and loss values. Provenance naturally represents data derivation relationships (i.e., transformations, parameter values, etc.), which provide important support in this analysis. Most existing provenance solutions define their own proprietary representations DL users choosing makes analysis interoperability difficult. We present Keras-Prov its extension, named Keras-Prov++, provides an analytical dashboard fine-tuning. Different from current mainstream solutions, captures applications using W3C PROV recommendation, allowing for help deciding on changing hyperparameters’ values after performance validation set. experimental evaluation Keras-Prov++ AlexNet real case study, DenseED, that acts surrogate model solving equations. During analysis, identify scenarios suggest reducing number epochs avoid unnecessary executions fine-tuning learning rate improve accuracy.