Advanced Computing in the Age of AI|Tuesday, September 29, 2020
  • Subscribe to EnterpriseAI Weekly Updates:  Subscribe by email

Getting Maximum Value from AI Deployments 

Artificial intelligence presents tremendous business opportunity, and organizations are increasingly investing in this technology to experience the value it can deliver. One survey of leading companies shows that the vast majority (91.6%) of global companies are increasing the pace of their AI investments, with 91.7% saying investment is “required to transform into agile and competitive businesses.”

However, the problem is that once created, as many as 87% of AI models don’t make it into production. And even when they do, they’re not always properly managed to deliver continuous value over time. To ensure organizations are deriving maximum ROI from their AI investments, they need MLOps.

MLOps, or Machine Learning Operations, is a combination of technology and practices that provide a scalable and governed way to deploy and manage machine learning models in production environments. With MLOps, companies can take AI projects from glorified science experiments to impactful, bottom line drivers.

Below, we explore just how MLOps can enhance the value of enterprise AI investments – from initial deployment and for years to come.

Reduce Time to Production

While data scientists can define the business problem they want AI to help solve, develop the model, and test it out, they generally need IT departments to deploy the model.

This is problematic for a few reasons. First, IT workers don’t have the same skill sets as data scientists –they’re backgrounds are in infrastructure, application monitoring, security and software development. In many instances, they have no idea what the model is or what it is meant to do. Second, circumstances change and models need to be updated even after they’ve made it to production – something IT isn’t able to do because they use different operating systems and programming languages than data scientists.

MLOps can help with this problem by creating a centralized hub where they can work together to get AI operational. As an automated process, MLOps operates on its own so data scientists and IT don’t need to worry about recoding languages, different operating systems or models drifting. Instead, it just seamlessly and efficiently puts the model into production so it can start to deliver business value.

Monitor and Govern Models

Once models make it into production, they need to be monitored and appropriately governed. Many companies don’t realize that AI models drift over time – meaning their predictions become less accurate if they aren’t updated frequently. Therefore, model monitoring is incredibly important to ensure they are leveraging the most accurate data and are performing properly. If not properly monitored, organizations could experience a loss of trust among business stake holders or even a loss in revenue.

Model governance is equally important to ensure businesses are best using AI. As companies develop models for critical customer-facing and business process applications, the need for governance becomes vital. The goal behind production model governance is to maximize the chance of a successful deployment and minimize risk through control of access and established update procedures. By limiting who can update models, maintain appropriate activity and prediction logs, and adequately test models, businesses can minimize risk, ensure legal and regulatory compliance, and create a repeatable process to scale AI adoption.

Prepare for the Unexpected

As we have all observed over the past few months, the world can change quickly – upending models and therefore appropriate business decisions and operations. While these unusual times can confuse models and make predicting things like consumer behavior harder, MLOps can help companies take these abnormal datasets and put them to use even in the most extreme and turbulent circumstances, such as the COVID-19 pandemic.

With MLOps, organizations can “reset” models by re-training them on newer datasets, and swiftly re-deploy them to production on the fly. They can give guidance on how, and by how much, data has drifted, underscoring which models are no longer making accurate predictions. There’s also historical data from other moments in time – such as the 1973 oil crisis or hurricanes – that can help businesses predict the long-term implications of a crisis as major as COVID-19. Since the impact of the virus shifts daily, MLOps can simultaneously keep tabs on where data might be skewed and use automation to be alerted of data and accuracy drift as soon as it happens. This then allows businesses to readjust their models quickly and get the insights they need to keep pace with the market.

When MLOps is used in conjunction with other AI tools, predictions become more accurate and businesses can adjust course swiftly to meet evolving consumer needs. For example, when models are trained on new datasets, like the ones mentioned above, and then applied to demand forecasting, brands can more accurately anticipate needs – ensuring that shortages on products such as hand sanitizers and toilet paper do not happen again. Likewise, when models are monitored and updated regularly with MLOps, companies can apply methods like time series to see lasting impacts on their business and use it to solve additional sets of problems such as optimizing staffing levels, managing inventory and more.

The world is changing by the day, and companies need accurate and trustworthy predictions to help them steer business decisions. However, until AI models are deployed and monitored at scale, enterprises will remain in the dark. By adopting to MLOps practices and solutions, enterprises can have the tools they need to navigate unprecedented times, while also finally seeing full value from their AI investments.

About the Author 

Sivan Metzger leads DataRobot’s MLOps & Governance Business, which helps companies finally obtain value from their ML initiatives by automating and scaling deployment, monitoring, management and governance of their ML models and applications in production. Sivan joined DataRobot via the acquisition of his company, ParallelM, where he served as CEO, and after spending two decades as an enterprise software business leader in many capacities at companies such as Kenshoo and Mercury Interactive.

Add a Comment

Do NOT follow this link or you will be banned from the site!
Share This