Improving our chances at achieving true business value through MLOps
This past June 3rd to 7th, ALTEN hosted its first International Tech Week, featuring a series of distinctive and insightful webinars on the given theme: Understanding the Impact of AI & ML on Our Industries and Daily Life. One standout session was Improving our chances at achieving true business value through MLOps, delivered by Jeroen Bleker, Machine Learning Consultant at ALTEN Netherlands.
With almost 8 years of experience in Data Science and having completed a PhD in Machine Learning for Radiology, Jeroen delves into MLOps, an emerging solution that aims to enhance the odds of achieving tangible business value by addressing these challenges. In this first article of a two-part series, he will touch base on some of the necessary background knowledge mentioned throughout his talk to enable a clearer comprehension of MLOps elements.
The hype around AI
The interest in Artificial Intelligence (AI) and its well-known subcategories, Machine Learning (ML) and Deep Learning (DL) (Fig. 1), has seen an almost exponential growth in the business sector in recent years. A significant catalyst for its current popularity was the introduction of ChatGPT (Generative AI), and the realisation that AI might be able to add substantial value. As Deborah Leff, CTO for Data Science and AI at IBM, stated at Transform.AI 2019:
“If your competitors are applying AI, and they’re gaining insights that allow them to accelerate, they’re going to pull ahead really, really quickly.”
Both this realisation and the widespread popularity of generative AI have seemingly led to a pervasive fear in the business sector of missing out on the added value through AI. This fear is perhaps not unwarranted, as there is no doubt that the proper implementation of AI can enhance operations, decision-making, and customer experiences.
Effective use
The effective use of AI can lead to significant competitive advantages. Routine tasks can be automated, and AI can be utilised to predict customer behaviour. Both examples should result in increased company efficiency. AI might be employed to innovate existing services and products or to develop new ones based on AI. Company revenue can be improved by optimising the existing situation using AI or creating new sources of revenue based on AI. Across various business functions, 63% of respondents to a McKinsey global AI survey reported revenue gains from AI adoption. High performers and front runners are nearly three times more likely to achieve revenue increases of more than 10% (McKinsey, 2019).
The effective use of AI stands or falls based on the approach and execution of projects. In practice, we see that more than 80% of ML projects do not make it to production (McKinsey, 2019). Machine Learning Operations (MLOps) has been introduced to address this challenge.
Machine Learning
Before delving into MLOps, it is important to understand its most important components: Machine Learning (ML) and Operations (Ops). Artificial Intelligence is a broad technology domain focused on performing tasks and simulating problem-solving and human-level intelligence. As previously stated, ML is a well-known subcategory of AI that deals with the creation of models (Fig. 1). Fundamentally, ML aims to perform tasks on new data using historical data. ML consists of four main types: Supervised Learning, Unsupervised Learning, Semi-supervised Learning, and Reinforcement Learning. The differences lie mainly in the approach to learning and the completeness of the input and output data. Figure 2 shows three examples of Supervised and Unsupervised Learning.
Operations
Aside from Machine Learning, it is important to understand what Operations entails. Over the last decade, there has been a swift increase in acronyms that are partly based on operations and use the ‘Ops’ suffix. The use of these acronyms started with development and company/IT operations, leading to DevOps. The goal of DevOps and other ‘Ops’ acronyms is to merge people, processes, and technology to deliver value optimally. In other words, ‘Ops’ acronym practices help companies better define processes, improve the quality of their output, and operate at a higher speed.
The 3 domains of MLOps
MLOps aims to streamline the end-to-end lifecycle of ML models. Figure 3 provides an overview of the three domains that collectively constitute MLOps: 1, 2, 3. Data Engineering is a domain that encompasses everything related to the collection, storage, and usage of data. A proper implementation of MLOps should lead to a faster time-to-value, reduced risks, and enhanced collaboration.
Implementation of MLOps
MLOps encompasses many specific elements and may continue to grow. Each element aims to solve a specific challenge that occurs during the ML lifecycle and can interfere with the effective use of AI. These elements can focus more on people and collaboration or on the technological process of ML. I believe that each MLOps element is best categorised based on their estimated importance for achieving effective use of AI/ML: Essential, Recommended, and Extra. Items within these priorities are further subdivided based on their occurrence in the end-to-end ML lifecycle. For some sectors, a small part of these priorities might be upgraded or downgraded based on sector requirements. This aligns with the fact that no optimal one-size-fits-all approach currently exists for MLOps.
A proper implementation of MLOps elements in a MLOps platform should account for element priorities and be flexible enough to support workflows while integrating with existing or forthcoming solutions. Other considerations when designing an MLOps platform include organisational skill set, level of tool abstraction, build vs. buy, and single element vs. full platform.
To delve even deeper into the implementation of MLOps elements, stay tuned for Jeroen’s 2nd article coming soon!
Write the story with us!