Evaluating the Effects of Convolutional Neural Network Committees
Fran Jurišić, Ivan Filković, Zoran Kalafatić · 2016
Many high performing deep learning models for image classification put their base models in a committee as a final step to gain competitive edge. In this paper we focus on that aspect, analyzing how committee size and makeup of models trained with different preprocessing methods impact final performance. Working with two datasets, representing both rigid and non- rigid object classification in German Traffic Sign Recognition Benchmark (GTSRB) and CIFAR- 10, and two preprocessing methods in addition to original images, we report performance improvements and compare them. Our experiments cover committees trained on just one dataset variation as well as hybrid ones, unreliability of small committees of low error models and performance metrics specific to the way committees are built. We point out some guidelines to predict committee behavior and good approaches to analyze their impact and limitations.