Computer Vision News - November 2018

Spearman rank correlation coefficients of the validation errors between different training times: A performance comparison between manually constructed architectures and the network architecture found by BOHB with a training time limit of 3 hours. All networks used the same hyperparameter space for optimization. You can see that by using simultaneous architecture and hyperparameter optimization the authors achieved competitive results: Conclusion: In this paper, the authors showed the advantages of simultaneously jointly optimizing architecture and hyperparameters. They demonstrated the powerful effect -- in term of training time -- of hyperparameter selection coupled with architecture selection. The authors demonstrated that this challenge can be solved with BOHB optimization method. Research 17 Research Computer Vision News