Self-Adaptive Machine Learning Architectures for Intelligent Workflow Optimization in Cloud Environments

Abstract

Adaptive Machine Learning Cloud Services are capable of adjusting their structure or behavior. They may, for example, add or remove components based on capacity demands or improve the quality of the service with more accurate algorithms. Furthermore, it is possible to adapt the structure of a Machine Learning architecture so that the inputs and outputs relate correctly for a particular task, such as classifying objects that have been detected in an image. Although there is significant literature describing Machine Learning as a Service, there is still little work that considers the specific needs of self-adaptive Machine Learning systems. In a cloud context, self-adaptation can take place in a service-oriented architecture and several factors can trigger this self-adaptation, including system or user needs. A guiding principle for resource management is elasticity, the ability to adapt the level of computing resources allocation to the current demand. Therefore, it should also be possible to manage the internal resources of a self-adaptive Machine Learning Service. A cornerstone concern is that the adjustment of the service requires monitoring. To avoid the overhead of monitoring particularly when resources are idle, a different approach is employed here, focusing on the reference behavior of a process. Instead of the intrinsic properties measured in several other monitoring papers, external indicators of quality are coupled to a feedback loop that triggers adjustment of the internal behavior.