Computer Vision News - December 2018

30 Computer Vision News Focus on For more explanations and examples, see the following websites: https://automl.github.io/HpBandSter and https://www.automl.org/blog_bohb/ Drawbacks: To use BOHB (also Hyperband) you must be able to set (and allocate) a meaningful budget: The low budget used for each training needs to (relatively) cheaply give a good indication of the function’s performance when the full budget will be used to run it on the entire dataset with a longer run-time. That is, the relative ranking of different hyperparameter configurations needs to be correlated with the relative ranking of the same configurations for the full budget. If the evaluations arrived at for low budgets are biased in some way, or simply too noisy, to be a good indication of the configurations that should be used for optimal performance on the full budget, then the Hyperband element of BOHB is a waste. In these cases, BOHB will just be k times slower than regular BO, where k is the number of Hyperband iterations implemented by that BOHB (this is the reason Hyperband can actually be worse then random search in the worst case scenario.) Feedback of the Month We got in touch with RSIP Vision with a challenging project with many adversities and imponderables. The good and successful cooperation with RSIP Vision allowed us to take a big step forward. We really much appreciated their project management and the clear and transparent way of communication ! Thank you very much! Christiane Michaelsen Managing Director, IDENTT SWISS GmbH Focus on: BOHB