AO, Highlight
Best paper Award IEEE International Conference on e-Science
Published on
Hyperparameter optimization is crucial for maximizing the performance of learning workflows, but it can be incredibly time-consuming.
Unlike traditional centralized methods with unique queue, decentralization empowers each worker to run on its own, but sharing their results globally.
The results? Not only did we see improvements in model accuracy, but we also witnessed faster convergence on a benchmark for personalised Cancer treatment. The optimal hyperparameters are discovered after only 25 mins when using 1,920 parallel workers on the Polaris supercomputer! Results which could not be achieved before even after several hours. This is a game-changer for large-scale hyperparameter (Bayesian) optimization!
Open-source : documentation and code available within the DeepHyper Package (https://deephyper.readthedocs.io).