Captura de dados de proveniência para apoiar a análise de hiperparâmetros em redes de aprendizado profundo

Abstract

Convolutional Neural Networks (CNN) training requires adjusting hyperparameters until a configuration is found in which the metric chosen by the specialist is satisfactory. For this, it is necessary to analyze CNN training data such as the hyperparameters configuration and their relationships with the data derivation. In this sense, provenance data can assist in this analysis by presenting metadata and the derivation of the training data. Current solutions for the analysis of hyperparameters configurations do not follow representation patterns of data derivation and do not allow analysis during training. Also, such approaches require CNN training to be executed under a portal or execution environment and/or have a significant impact on CNN’s training time. In this dissertation, we present CNNProv, a provenance data capture solution, which allows the analysis of hyperparameters values during training. CNNProv adopts the W3C PROV standard to represent provenance data and can be used as services, regardless of a portal, thus contributing to the CNNs training phase. The experiments conducted with the training of CNNs in different scenarios use a classic CNN, AlexNet, and a seismic imaging application. The results show the suitability of CNNProv for the hyperparameter analysis with a negligible overhead of up to 4%.