Blog series: Becoming an MLOps Expert (1)
Adoption of Artificial Intelligence (AI) is increasing year-on-year. Within this, research shows that Machine Learning (ML) applications and platforms account for almost two-thirds of all AI investment¹. For those with the skills, tools and understanding, ML represents a significant market advantage that will increase over time, as leaders find new applications for the technology, propelling them further ahead of their peers. In fact, McKinsey suggests current AI investment dynamics “could create a wider divide between AI leaders and the majority of companies².” Put simply, inaction is not an option. Organisations must take advantage of ML now, or their competitors will. Yet while pockets of ML research and experimentation can be found across several industries, only 13% of ML projects make it into production³. Why? Enter MLOps. In this blog, the first in a four-part series focusing on MLOps (Machine Learning Operations), I will discuss the challenges around current approaches to ML and the forces underpinning the shift to MLOps.
Why data science leaders are embracing ML as the next frontier
ML enables more complex analysis of data than traditional analytics, helping to identify patterns that may otherwise go unnoticed. Based on algorithms that detect patterns; ML learns how to make predictions and recommendations by processing data, rather than receiving explicit programming instructions, like traditional analytics. This means organisations no longer need thousands of analysts to come up with all potential situations – the algorithm can imagine these scenarios for you. Instead of working with three-four data features or dimensions, you can work with hundreds. ML algorithms also adapt in response to new data, to improve predictions and recommendations over time. Today’s dynamic digital environment is supportive of ML. Large volumes of rich data are being generated that can be processed and analysed at the click of a button. Data scientists also have access to more capable and usable ML models and methods, coupled with the reduced cost of compute/processing through the cloud. These factors allow ML to be effectively applied to more organisations through more use cases – from risk management and fraud defection to financial trading, forecasting and business process automation. However, many projects are doomed to fail.
Key challenges to ML in a non-operationalised world
ML projects often result in low success rates because teams lack the right people, processes and technologies to make them work in production. Generally, ML projects lack rigour and standardisation, making it impossible to repeat, audit or scale in any meaningful way. Data is often poorly managed between the experimentation phase and live. For example, changes in data formats or non-functional requirements used to select and train the model may undermine live deployment. For example, a high demand consumer service that requires quick response times may require a level of scalability that was not tested in the experimentation phase. Other non-monitored data issues can result in mis-predictions, false positives and the need for retraining – a laborious and time-consuming task. Monitoring of the ML model is also a significant new area to cover. If a model does break, it needs to be detected and fixed quickly. This requires constant monitoring, which when conducted manually is a major drain on resources. This makes scaling ML projects more difficult.
How to make the shift to MLOps
For ML to reach its full potential and to increase the rate of live ML projects, it needs to be operationalised. By the end of 2024, Gartner predicts 75% of organisations will shift to operationalising AI. MLOps is the primary process that will underpin this shift to operationalising AI. Machine Learning Operationalisation, or MLOps, holds the key to success. It is a set of practices that combines ML, DevOps and Data Engineering, aiming to deploy and maintain ML systems in production reliably and efficiently. MLOps should apply across all ML projects with standardised steps and repeatable tools/software. It is not about one model or one data scientist, but about multiple projects and multiple people getting the ML “job” done. As such, MLOps requires a holistic assessment of the development process, covering the whole ML process from discovery, piloting and deploying to monitoring and operations. Data scientists must begin to operate ML processes with a rigorous, structured and repeatable process at a greater scale and speed. This process needs to be 100% transparent to both support moving the ML model to live and ensure its explainability.
Tune in next week…
Hopefully, this has given you an understanding of why you need to put the “ops” into your ML projects. In my next blog, we will discuss the benefits of MLOps further, outlining specific use cases that can help drive value to your business.