Take your AI models out of the lab and into the market. That’s the challenge. But how do you deploy in days, not months, while ensuring they deliver expected results—and remain fully compliant, even as AI regulations advance?
That’s the gist of a new whitepaper by Kevin Petrie, vice president of research at Eckerson Group. Petrie, who launched data services businesses for EMC Pivotal, now serves as an industry analyst for Eckerson Group. In his newest whitepaper, Deep Dive on Machine Learning Platforms, Petrie shines a light on artificial intelligence and machine learning (AI/ML) models.
“To reduce risk and deliver results, enterprises need to build and manage multiple ML models in a controlled way,” he notes. This requires models that deliver value, of course, and are also built and managed in a controlled way.
The need to deliver results while reducing risk is a central theme. For Petrie, this demands an open collaboration across business units, data scientists, data engineers, and ML engineers. Most enterprises, however, can’t achieve this using homegrown tools. The answer: ModelOps.
The OS for Your AI
ModelOps is essential for all predictive and people-facing algorithms, including credit fraud, insurance risk assessment, dynamic pricing, medical decisions, intrusion detection, and more. In Deep Dive on Machine Learning Platforms, analyst Kevin Petrie reviews the top ModelOps tools and assesses their benefits and use cases:
- Unlocking the value of AI/ML software
- The need to build and manage multiple AI/ML models in a controlled way
- Increasing concerns over regulatory compliance
- Reviews of major ModelOps platforms, including top features and differentiators
- Which ModelOps platforms are optimized for on-premise, cloud, or hybrid environments
- How ModelOps platforms fit into a larger data-driven architecture
This highly visual whitepaper reveals the push to accelerate time-to-model pipeline deployment across staging, development, test, review, and production. Plus, how ModelOps anticipates risks, prevents bias, and keeps your organization compliant with advancing AI regulations.
A Machine Learning Platform Architecture
“ML models can run amok in many ways,” Petrie says. They may focus on the wrong patterns, make bad predictions, and fail to spot business changes or shifts in the marketplace. ModelOps solutions are designed to limit or eradicate this.
Petrie outlines a proposed machine learning platform architecture. This platform can provide an “environment in which data science teams can manage machine learning in an efficient, governed and scalable way.”
Petrie then lists the three stages of the machine learning lifecycle:
- Data and feature engineering. In this stage, a data scientist works with a data engineer to transform input data and label the historical outcomes.
- Model development. Here, data scientists “train” the algorithm. They must check results and train it until that algorithm delivers the desired accuracy.
- Model production. Finally, the ML engineer implements the model into production workflows.
Petrie also recommends that an ML platform include visualization tools: “Data science teams visualize data during each stage of the life cycle—for example, to explore data, define features, or inspect the outputs of training or production models.”
A ModelOps solution that allows for a shared library of ML algorithms, current model components, and models in development is also recommended.
Reviews of ModelOps Solutions
Finally, Petrie provides reviews of the existing leading ModelOps solutions. He outlines the strengths and weaknesses of these tools—and the companies that developed them.
Reviewing TIBCO’s ModelOps solution, Petrie states: “TIBCO offers an ML platform on top of its hybrid data platform. TIBCO streams and virtualizes data on hybrid infrastructure; and helps manage data and feature engineering, model development, and model production. It also assists the lifecycle with ML ecosystem support.”
Petrie also notes the TIBCO solutions build on TIBCO’s “heritage in BI visualization” and “encourages business owners, citizen data scientists, and other stakeholders to actively contribute to the ML lifecycle for rapid real-world value and ROI.” Petrie also notes how TIBCO ModelOps works closely with TIBCO Spotfire and TIBCO Streaming.
To read the in-depth review and the entire whitepaper, click here.