Extending the hypergradient descent technique to reduce the time of optimal solution achieved in hyperparameter optimization algorithms

There have been many applications for machine learning algorithms in different fields. The importance of hyperparameters for machine learning algorithms is their control over the behaviors of training algorithms and their crucial impact on the performance of machine learning models. Tuning hyperparameters crucially affects the performance of machine learning algorithms, and future advances in this area mainly depend on well-tuned hyperparameters. Nevertheless, the high computational cost involved in evaluating the algorithms in large datasets or complicated models is a significant limitation that causes inefficiency of the tuning process. Besides, increased online applications of machine learning approaches have led to the requirement of producing good answers in less time. The present study first presents a novel classification of hyperparameter types based on their types to create high-quality solutions quickly. Then, based on this classification and using the hypergradient technique, some hyperparameters of deep learning algorithms are adjusted during the training process to decrease the search space and discover the optimal values of the hyperparameters. This method just needs only the parameters of the previous two steps and the gradient of the previous step. Finally, the proposed method is combined with other techniques in hyperparameter optimization, and the results are reviewed in two case studies. As confirmed by experimental results, the performance of the algorithms with the proposed method have been increased 36.62% and 23.16% (based on the best average accuracy) for Cifar10 and Cifar100 dataset respectively in early stages while the final produced answers with this method are equal to or better than the algorithms without it. Therefore, this method can be combined with hyperparameter optimization algorithms in order to improve their performance and make them more appropriate for online use by just using the parameters of the previous two steps and the gradient of the previous step.