建築工学技術

抽象的な

Neural Network Convolution Architecture for Image Classification: Fitness Landscape Analysis

Icroh Meattini

It is unclear which hyper parameter search technique will be most successful because the global structure of the hyper parameter spaces of neural networks is not well understood. In order to offer guidance on suitable search methods for these spaces, we study the topographies of convolutional neural network architectural search spaces in this research. We investigate the overall structure of these spaces using a traditional method (fitness distance correlation) and a more contemporary instrument (local optima networks).