Productionize your Machine Learning Models
Predictive Analytics is a two-stage process. In the first stage, machine-learning algorithms are used to build models from historical data. In the next stage, the predictions from these models make or influence a business decision in the production systems.
Moving the prediction models to production is the primary roadblock that stands in the way of getting value out of machine learning models. While it is relatively straightforward for data scientists to build powerful models in environments like Spark, SAS using open source tools like Conda and TensorFlow in Python and R, the business value comes only from deploying the machine learning models into production. IT professionals play a critical role in ensuring that models developed by their business peers and data scientists are suitably deployed so that they succeed in serving predictions that optimize the business processes. Real time production systems are often in a completely different tech stack as compared to the model development environment. There is no single method for deploying AI models that works in every situation and caters to every use case. Most of the methods are tied closely to a specific infrastructure, have a long deployment cycle from development to production, or are not suitable for a real time production system at all.
We will discuss different methods to deploy the prediction models in a production environment. An analysis of the pros and cons of each technique will be discussed and a technical guidance for choosing the right solution based on the type of data, business needs and environmental support will be provided. The following techniques of deployment would be covered:
- Re-writing of the model code in the language of the data engineers
- Hosting models in a web service and exposing prediction endpoints
- Serializing and De-serializing the models using different libraries
- Standards-based approach eg. PMML
Outline/Structure of the Talk
This work will serve as a reference for adopting the right methodology for deploying machine-learning models in hybrid environments (Cloud and on-premise). It will share insights to all stakeholders on the tools and technologies, processes, pitfalls, industry best practices and standards in various scenarios of model deployment. The deployment challenges of the machine learning models in production for any business will thus be effectively addressed.
Data Engineers, Data Scientists
Prerequisites for Attendees
Basic understanding of developing and deploying Machine Learning models