부산대학교 도서관

Although Machine Learning and Deep Learning have advanced significantly in their ability to perform supervised, unsupervised and reinforcement learning tasks quite well, integrating them into applications for which they are meant to provide intelligence is not as seamless as it should be. There are many reasons for this - primary being that the kind of skills needed to understand the business need, understand the data, develop models, develop applications, and integrate models with applications are different and it is very difficult for one person or just a group of application developers or just a group of data scientists to have all these skills. This makes it challenging to productionalize the developed ML models fast or at all in order to experiment with them further. This paper demonstrates how certain components of the Acumos AI platform project can be used to take models developed using H2o, Java, Spark to production by deploying them as microservices, automatically. The same concept can be (and has been extended by Acumos) to Python, R and ONNX models. What this enables is - multiple heterogeneous models written by different developers or different teams or different organizations in different languages and frameworks becoming functioning microservices that would provide intelligent APIs to the business application in question. These models can then be easily shared with different individuals and organizations and operationalized easily. Using Acumos, these microservices can be deployed such that they also communicate and co-ordinate with each other to do much more complex tasks. We will talk about the 5G Network Slicing usecase, create an ML model for 5G Network Slicing and use certain Acumos components to make it shareable and operationalize it as a predicting microservice.