Research On The Universal Bi-level Optimization Calculation Tool For Meta-learning

In recent years,meta-learning and its related applications have gradually become a hot research topic.It improves data efficiency,transfers prior knowledge and saves computing resources through training model to learn to learn.At the same time,many different types of meta-learning computing strategies have been proposed continuously,so it is particularly important to model the meta-learning problem systematically and analyze different types of methods.In this paper,the optimal control theory and dynamic system are introduced to help remodel the mathematical model of meta-learning problems,a universal bi-level optimization framework is proposed to summarize mainstream meta-learning methods based on optimization,and a general calculation tool named BOML is designed and implemented to solve the related application problems of meta-learning.BOML has the characteristics of modularity,easy expansion and strong generality.It can quickly construct the bi-level optimization model through simple interfaces,and combines different bi-level optimization strategies to solve the meta-learning problem.BOML can design new meta-learning algorithms through different combinations of upper and lower level strategies.Furthermore,the calculation tool is accessible on Git Hub and provides online documentation for users to read and further update,which provides a complete set of supporting resources for researchers.In the experimental part,BOML not only considers few shot classification and hyper cleaning experiments,but also considered in combination with other areas of knowledge,such as spectral normalization,knowledge distillation strategy to help algorithms improve ascension in the orthogonality and other indicators of performance.Based on the research of the universal bi-level optimization calculation tool,this paper successfully remodels the meta-learning problem,and combs a series of optimization-based meta learning solutions,so as to provide a new research angle for researchers,and promote the research of new meta-learning methods and application fields.