This post first appeared on Federal News Network. Read the original article.
The Centers for Disease Control and Prevention is leveraging machine learning to predict the spread of COVID-19. The Postal Service is using the technology to speed package delivery. The Transportation Department is piloting use of ML to forecast structural safety of highway bridges. The Defense Department is testing it to visualize terrain and “see” around obstacles.
In fact, a growing number of federal agencies are exploring ML and other forms of artificial intelligence to further their missions or improve services to citizens. As they do, they need proven ways to ensure they’re implementing ML effectively and achieving positive returns on their ML investments.
Enter Machine Learning Operations (MLOps), a set of practices intended to optimize the reliability and efficiency of ML design, development and execution. Influenced by the widely adopted DevOps approach to creating and operating custom applications, MLOps combines methodologies and technologies to accelerate deployment of ML models. There are great tools from software vendors and the open source community available today such as cnvrg.io, c3.ai, Databricks and SAS; with open source tools such KubeFlow, MetaFlow, Kedro and MLFlow.
MLOps is already proving its value in government. In fact, documenting and enforcing MLOps makes agencies twice as likely to achieve AI goals and three times more likely to be prepared for AI risks. That’s according to Deloitte analysis of government respondents to its 2021 State of AI in the Enterprise survey.
MLOps is used by data scientists and other AI experts. But what makes it particularly appealing is that it can allow agencies with less specialized technical skills to benefit from ML models. The top benefits are rapid innovation, creating reproducible models and workflows, creating management structure and flow for the entire machine learning lifecycle, all focusing on sharing, learning and fastest time to answers.
Meeting ML challenges
ML involves complicated mathematics. Training and maintaining ML models can be complex and time-consuming. Agencies sometimes struggle to attract and retain experienced data scientists or get budget approvals for lengthy, costly ML pilots.
MLOps addresses these issues. The approach can help simplify, streamline and even automate portions of ML development and operations. It can reduce the number of ML experts required and the time and cost to progress from business requirement to working solution.
MLOps provides an iterative process for designing, developing and operating ML models. It starts with a standardized approach to determining the business requirements of the ML use case. It then streamlines identification of the necessary data inputs, appropriate ML algorithm and requirements of the ML model. Finally, it automates the feeding of outputs back into the model for continuous improvement, replacing the time-consuming, manual approach of repeatedly training and testing the model.
Many agencies will be able to implement MLOps without investing in new hardware or other infrastructure. But some technology tools can help. For example, if you want to leverage large datasets in your ML use case, you might need hardware that supports persistent memory. If you’re implementing a lot of ML, you might benefit from MLOps orchestration software to manage your end-to-end ML pipeline.
But a growing number of data analysis and visualization products have MLOps capabilities built in. Some can even walk you through the process from identifying business requirements, to specifying the dataset, to selecting the ML algorithm, to creating the ML model.
In addition, major providers of public cloud services offer suites of tools that include ML algorithms. You can use MLOps methodologies to identify the algorithm for your use case and test an ML model to see if it meets your needs.
Following MLOps best practices
MLOps specifies ML best practices. But there are also best practices for getting the most out of MLOps. That always starts with getting a clear picture of the problem you want to solve, whether that’s identifying structural problems in critical infrastructure or flagging interactions among prescription medicines.
Next, review your inventory of existing software to see if you already have tools that offer MLOps capabilities. Your data management software or public cloud service might provide a library of relevant ML datasets or algorithms. So might your department or partner agencies. The DoD, for instance, is building marketplaces of ML algorithms and models. You can take advantage of these resources to accelerate your ML deployment.
Don’t neglect cybersecurity. If you’re working with sensitive data such as personal identifiable information (PII) or classified intelligence, you’ll want hardware that can encrypt ML data both while it’s at rest in a database and while it’s in use in computer memory.
Finally, take into consideration the potential for bias in datasets and ML models. There are documented cases where facial-analysis programs offered by leading IT providers were biased based on gender and skin color. As a result, their outputs could be wildly inaccurate. Using a biased dataset or model from an ML library only perpetuates these errors. The good news is that the documentation and enforceable processes of MLOps help ensure more transparent and trustworthy AI.
From planning supply networks to making public health recommendation, from predicting maintenance for critical infrastructure to isolating anomalies in cyberwarfare, ML will find more use cases among more federal agencies. MLOps can help your agency establish best practices for efficiently and cost-effectively designing, developing, testing, deploying and maintaining ML solutions. The result will be ML outcomes that better serve citizens and support your mission.
Gretchen Stewart is chief data scientist for Intel Public Sector.