Algorithms for speeding-up the deep neural networks for detecting plant disease

In designing an artificial network, different parameters such as activation functions, hyper-parameters, etc. are considered. Dealing with large number of parameters and also the functions that are expensive for evalualtion are very hard tasks. In this case, it is logical to find methods that results in smaller number of evaluations and improvements in performance. There are various techniques for multiobjective Bayesian optimization in deep learning structure. S-metric selection efficient global optimization (SMS-EGO) and DIRECT are one of the many techniques for multiobjective Bayesian optimization. In this paper, SMS-EGO and DIRECT techniques are applied to deep learning model and the average number of evaluations of each objective including time and error are investigated. For training and validating the deep network, a number of images present various diseases in leaves are provided from Plant Village data set. The simulation results show that by using SMSEGO technique, performance is improved and average time per iteration is faster.