Computer Vision News - January 2019

Repeat training using different methods, different configurations and various architectures, comparing the performance of networks trained from random initialization with those pre-trained then fine-tuned -- shows it cannot be just chance that the overall data needed proves equal again and again, whether starting from scratch or using pre-trained networks. The methods are equivalent. Another experiment ran by the authors consisted in training the pre-trained network to find the optimal hyperparameter settings. The authors used the hyperparameter settings discovered by training the pre-trained network to train their random-initialization network from scratch -- and achieved equal results using just one third of the data. The figure below presents training accuracy (purple shows random initialization, grey shows pre-trained then fine-tuned). Research 8 Research Computer Vision News