A Study On Meta Learning Optimization Techniques

In the last decade, the domain of Artificial Intelligence (AI) has undergone rapid advancement and has achieved ability to mimic near-human intelligence in multiple aspects. Deep Neural Networks (DNNs), when trained with large datasets and optimal computational resources can achieve great outcomes, but the learning process is exhaustively time consuming. Along with this there still exists several challenges with respect to computational resources constraints, scarcity of data related to specific task and so on. In various situations, where the data is scarce, the algorithms are incapable of learning tasks quickly and accurately similar to that human intelligence. Meanwhile, the problem has been addressed with advancements in deep meta-learning in multiple researches. Meta-Learning has wide range of applications where in the insights from meta-data of the either data, tasks or the previously trained models can be utilised to optimise the learning process. To get an in-depth overview of available meta-learning techniques for DNN model optimization, we have carried-out a survey of optimization-based meta-learning techniques, their advantages, drawbacks and open challenges.