Suppose your organization’s data science teams have documented business goals for areas where analytics and machine learning models can have business impact. Now they are ready to start. They tagged datasets, selected machine learning technologies, and established a process for developing machine learning models. You have access to a scalable cloud infrastructure. Is that enough to give the team the green light to develop machine learning models and deploy the successful ones to production?
Not so fast, say some machine learning and artificial intelligence experts who know that every innovation and production deployment comes with risks that require reviews and remediation strategies. They advocate establishing risk management practices early in the development and data science process. “In data science or any other similarly oriented business, innovation and risk management are two sides of the same coin,” said John Wheeler, Senior Advisor of Risk and Technology at AuditBoard.
Similar to application development, software developers don’t just develop code and deploy it to production without considering risks and best practices. Most companies establish a Software Development Life Cycle (SDLC), shift devsecops practices, and create observability standards to address risk. These practices also ensure that development teams can maintain and improve code once it’s deployed to production.
The equivalent of SDLC in machine learning model management is modelops, a set of practices for managing the lifecycle of machine learning models. Modelops practices include how data scientists build, test, and deploy machine learning models to production, and then monitor and refine ML models to ensure they deliver expected results.
Risk management is a broad category of potential problems and how to fix them, so in this article I’ll focus on those associated with modelops and the machine learning lifecycle. Other related risk management topics are data quality, data protection and data security. Data scientists also need to check training data for bias and consider other important responsible AI and ethical AI factors.
Speaking to several experts, below are five problematic areas that Modelops practices and technologies can help address.
Risk 1. Development of models without a risk management strategy
In the 2022 State of Modelops Report, more than 60% of AI business leaders said risk management and regulatory compliance are challenges. Data scientists are generally not risk management experts, and in organizations, a first step should be to work with risk management executives and develop a strategy that aligns with the Modelops lifecycle.
Wheeler says, “The goal of innovation is to look for better ways to achieve a desired business outcome. For data scientists, this often means creating new data models to make better decisions. However, without risk management, the desired business outcome can come at a high cost. In striving to innovate, data scientists must also seek to create reliable and valid data models by understanding and mitigating the risks inherent in the data.”
Two white papers to learn more about model risk management are from Domino and ModelOp. Data scientists should also adopt data observability practices.
Risk 2. Increased maintenance through duplicate and domain-specific models
Data science teams should also create standards for which business problems to focus on and how to generalize models that work across one or more business domains and divisions. Data science teams should avoid creating and maintaining multiple models that solve similar problems; You need efficient techniques to train models in new business areas.
Srikumar Ramanathan, Chief Solutions Officer at Mphasis, recognizes this challenge and its implications. “Each time the domain changes, the ML models are trained from scratch, even using standard machine learning principles,” he says.
Ramanathan offers this remedy. “Through incremental learning, where we continuously use the input data to extend the model, we can train the model for the new domains with fewer resources.”
Incremental learning is a technique for continuously training models with new data or at a defined cadence. There are examples of incremental learning on AWS SageMaker, Azure Cognitive Search, Matlab, and Python River.
Risk 3. Deploying too many models for the capacity of the data science team
The challenge of maintaining models goes beyond the steps of retraining them or implementing incremental learning. Kjell Carlsson, Head of Data Science Strategy and Evangelism at Domino Data Lab, says, “A growing but largely overlooked risk lies in the ever-decreasing ability of data science teams to re-engineer and re-deploy their models.”
Similar to how development teams measure cycle time for deployment and feature deployment, data scientists can measure their model velocity.
Explaining the risk, Carlsson says: “Model speeds are typically well below demand, leading to a growing backlog of underperforming models. As these models become more important and anchored across companies – in combination with accelerating changes in customer and market behavior – a ticking time bomb is emerging.”
Dare I call this problem “model guilt”? As Carlsson suggests, measuring model speed and the business impact of underperforming models is the most important starting point for managing this risk.
Data science teams should consider centralizing a model catalog or registry so team members know the scope of existing models, their status in the ML model lifecycle, and who is responsible for maintaining them. Model catalog and registration capabilities can be found in data catalog platforms, ML development tools, and both MLops and Modelops technologies.
Risk 4. Getting bottlenecked by bureaucratic review boards
Let’s assume the data science team followed the organization’s standards and best practices for data and model governance. Are they finally ready to deploy a model?
Risk management organizations may wish to establish review boards to ensure data science teams are mitigating all appropriate risks. Risk reviews can be useful when data science teams are just beginning to deploy machine learning models in production and adopt risk management practices. But when is a review board necessary and what to do when the board becomes a bottleneck?
Chris Luiz, Director of Solutions and Success at Monitaur, offers an alternative approach. “A better solution than a top-down, post-hoc, and draconian executive review board is a combination of sound governance principles, software products that fit the data science lifecycle, and strong stakeholder alignment across the governance landscape. Process.”
Luiz has several recommendations on Modelops technologies. He says, “The tools must fit seamlessly into the data science lifecycle, maintain (and preferably increase) the speed of innovation, meet stakeholder needs, and provide a self-service experience for non-technical stakeholders.”
Modelops technologies with risk management capabilities include platforms from Datatron, Domino, Fiddler, MathWorks, ModelOp, Monitaur, RapidMiner, SAS, and TIBCO Software.
Risk 5. Failure to monitor models for data drift and operational issues
If a tree falls in the forest, will anyone notice? We know code needs to be maintained to support framework, library and infrastructure upgrades. Do monitors and trend reports alert data science teams when an ML model is underperforming?
“Any AI/ML model put into production is guaranteed to degrade over time due to the changing data of dynamic business environments,” said Hillary Ashton, executive vice president and chief product officer at Teradata.
Ashton recommends, “Once in production, data scientists can use modelops to automatically detect when models are starting to degrade (reactively via concept drift) or likely to start degrading (proactively via data drift and data quality drift). You may be notified to investigate and take action, e.g. B. retrain (update model), retire (requires complete redesign), or ignore (false alarm). In the event of retraining, remediation can be fully automated.”
What you should take away from this review is that data scientist teams should define their Modelops lifecycle and develop a risk management strategy for key steps. Data science teams should work with their compliance and risk officers and use tools and automation to centralize a model catalog, improve model speed, and reduce the impact of data drift.
Copyright © 2022 IDG Communications, Inc.