Developing machine learning (ML) models is a costly and timely business. Gartner says most organizations take nine months to integrate ML models from prototyping into production.
The research firm explains that there is much focus on developing analytical and ML artifacts. Still, many overlook the quintessential portion of operationalization to ensure the continuous delivery and integration of ML models within enterprise applications and business workflows.
In Asia, operationalizing ML models is also challenging. Only half of all AI proof of concepts (PoCs) make it to production, according to Amaresh Tripathy, global analytics leader from Genpact, a professional services company that focuses on digital transformation.
He notes due to the slower adoption of AI/ML, many enterprises in the region have yet to be motivated to scale and operationalize ML.
“Many teams are just beginning to unlock the power of these technologies through proof of concepts and experimentations,” says Tripathy. “The upcoming years will be pivotal in scaling AI/ML solutions for most leading organizations in Asia. But to succeed, they will need a reliable framework such as machine learning operations (MLOps).”
Accelerate innovation with MLOps
Derived from the core principles and benefits of DevOps, MLOps is a framework aiming to standardize the deployment and management of ML models.
According to Gartner’s report Understanding MLOps to Operationalize Machine Learning Projects, successful MLOps align with the continuous integration and continuous deployment (CI/CD) pipeline in the DevOps practice, allowing enterprises to speed up ML models from PoCs into production.
“Enterprises derive significant benefits of agility and repeatability, which creates a massive competitive advantage due to your ability to learn faster than the competition,” adds Manjunath Bhat, vice president analyst at Gartner.
He says MLOps also simplify some of the integration challenges of ML models governance, monitoring, and deployment. This reduces time to market and improves ROI on AI/ML initiatives.
“MLOps can help enterprises in Asia industrialize AI engineering practices across organizations by enabling the creation of consistent and coherent data pipelines,” says Bhat.
By adhering to different standards, security, ethics, and regulatory requirements on the MLOps, enterprises can build more accurate, responsible, and explainable models.
The roadblocks of MLOps
By creating a common architecture, MLOps helps operationalize data science and ML pipelines. Thus, Gartner identified MLOps as one of the major trends for 2021.
Yet, scaling and operationalizing ML models in Asia remains slow.
“Organizations struggle with scaling AI for many reasons,” notes Gartner’s report. “Security and privacy concerns, integration complexity and potential risks and liabilities exist on top of data challenges.”
On top of these concerns, the report states there is still a lack of understanding of AI’s benefits and uses. The unavailability of technology knowledge to operationalize AI also makes it challenging for enterprises to adopt MLOps.
“The root cause (of slow MLOps adoption in Asia) is a misbelief that AI/ML could outright eliminate the workforce driving USD900 billion worth of transactional tasks by automating approximately 50% of jobs like data collection, data processing, and professional services,” adds Tripathy from Genpact.
In addition to the lack of technical know-how, Bhat from Gartner says the organizational silo between different roles across the ML development life cycle also contributes to the slow adoption.
“Collaboration between development, data, security, and IT operations is one of the foundational prerequisites (for MLOps),” he says. “We see a lack of mature automation and the need for a cross-functional, multi-disciplinary team culture as one of the key barriers to adoption of MLOps in Asia.”
Early successes in Asia
Nonetheless, some early adopters in Asia are making headway in adopting MLOps.
Tripathy says one of them is Indonesia-based Gojek, an on-demand multi-services platform and digital payment group. To support the company’s super app and continuous development of more than 20 services, Gojek added Merlin to its ML platform. Merlin provides model management and deployment and model serving and monitoring function.
“The platform aims to enable rapid, scalable, and self-service model deployment by abstracting infrastructure complexity and autoscaling,” notes Tripathy.
Another example is an insurance company using MLOps to build more accurate and predictive models of different claim expenses. According to the MLOps playbook — developed by Genpact, National Association of Software and Service Companies (NASSCOM), and EY — the insurance company riding on MLOps can better manage its financial status and build better pricing models.
“To help organizations maximize the value of their AI and machine learning investments, we developed an MLOps playbook that sets the foundation for scaling these initiatives,” adds Tripathy. “In essence, our playbook enables teams in Asia to circumvent challenges that other industry leaders have already experienced.”
In addition to Genpact, Gartner also published reports on Demystifying XOps, to help data and analytics professionals leverage DevOps to operationalize data analytics and AI architectures.
“Most organizations recognize the value of automation and operationalization to the successful delivery of analytic and AI initiatives, but few do it well,” notes Gartner. “MLOps extend the software development techniques of CI and CD toward analytic and AI platforms. This is an entirely new way of working, but the transition is not as arduous as it may seem.”
Sheila Lam is the contributing editor of CDOTrends. Covering IT for 20 years as a journalist, she has witnessed the emergence, hype, and maturity of different technologies but is always excited about what's next. You can reach her at [email protected].
Image credit: iStockphoto/Tetiana Lazunova