| Summary: | Stochastic optimization in machine learning adjusts hyperparameters to reduce cost, which shows difference between actual value of estimated parameter and things predicted by machine learning model. Learning rate regulates the amount of alteration to a model in terms of the predicted error. Tuning learning rate is a challenging task. A too large learning rate affects the performance of a model while a too small learning rate may never converge at optimal or near-optimal solution. Adaptive learning rate adjusts learning rates based on performance of a model. This paper presents a survey on first-order stochastic optimization algorithms, which are the main choice for machine learning due to their speed across large datasets and simplicity. Stochastic gradient descent method, its variants and mini-batch algorithms will be elaborated. Adaptive learning rate embedded in the stochastic optimization algorithm can be further improved. This paper discusses learning rate adaptation schemes and how these affect stabilization in the value of learning rate which helps stochastic gradient descent to show fast convergence and a high success rate. This paper aims to offer useful insights towards the development of future stochastic optimization algorithms in machine learning. © 2023 Author(s).
|
|---|