The missing people in MLOps

MLOps — from the lab to the real world

Every month, hundreds of new articles are published in scientific journals that show proof-of-concept AI systems for a range of applications including: automated cancer detection, including at-home tests; automated fall detection systems that allow the elderly and vulnerable to maintain their independence but still be safe; and identifying early signs of diabetes from images that can be taken as part of your regular eye test. But how many of these are actually being used to help improve peoples’ lives? Very few in fact. (See an article by my colleague Bill Shepherd for a discussion sparked by a recent Nature Machine Learning paper that discussed just this.)

So why aren’t these amazing solutions to serious problems being used? Well, there is a big gulf between developing an AI system and operationalising or productionising it in a safe, reliable and usable way. This is where Machine Learning Operations (MLOps) comes in.

MLOps brings together scientists who develop these early stage AI systems with engineers and operations experts who are able to convert these AI models into reliable and scalable products such as apps on your phone or smart microscopes in the hospital.

A recent Towards Data Science article states that:

MLOps enables collaboration across diverse users (such as Data Scientists, Data Engineers, Business Analysts and ITOps) on ML operations and enables a data-driven continuous optimization of ML operations’ impact or ROI (Return on Investment) to business applications.

And if you read around the topic you might be lucky enough to see such complete MLOps pipelines like the one in this blog post arguing we need to ensure that data discovery and engineering isn’t left behind as MLOps develops:

What’s missing in MLOps? The experts!

Excellent! We’ve brought together data scientists and engineers so we can now productionalise our AI — right? Well, this is where things get complicated. Current MLOps pipelines largely come from domains where the people described as experts that the AI is trained to mimic are largely everyday people — shoppers or car drivers, for example. In these cases, the data scientists, AI scientists and even citizen scientists are able to ‘annotate’ large data sets with the features of interest.

But what happens in domains like medicine? Do we want our medical data being annotated by data scientists with no medical training? I think not. In fact, as different regions of the world start to develop regulation for AI systems many of them are beginning to question who annotates this training data, especially in domains that require a high level of expertise. This is most clearly represented by the EU Ethics Guidelines for Trustworthy AI, which explicitly demands that there be accountability for AI systems and their outcomes especially in critical applications. This is achieved through auditability of data and design processes in the development of the AI product, and this includes information on who annotated data and how data is annotated.

MLOps is a new and emerging field and products are now being released that satisfy the traditional MLOps pipeline of AI engineers, operations engineers and even data engineers. However, at gliff.ai, we believe that there are two main groups of people “missed out” by these tools: the domain experts and the regulatory experts.

Enter the experts — collaborative MLOps

First, let’s discuss who these experts are and why we need to include them when developing AI products.

Domain experts are those humans with the expertise we hope to mimic with our AI, i.e. surgeons, physiotherapists, pediatric nurses, marine scientists, aerospace engineers, and so on. As mentioned earlier, many of the AI products that are currently on the market mimic the skills of everyday people: shoppers, drivers, etc.. But in many domains where AI could fundamentally change lives, such as medicine, this expertise is only found in a few, select, highly-qualified, time-pressured and costly experts, such as clinical doctors. If we want to train an AI to do automated diagnosis, then we need to train our AI on data annotated by doctors specialised in that field — the domain experts.

Many existing MLOps products are command line or programming tools that require significant programming expertise that many domain experts, such as clinical doctors, don’t have. gliff.ai is aiming to combat this by creating a browser-based MLOps pipeline that aims to engage domain experts at all stages of the MLOps process. This isn’t about doing AI without data engineers and AI scientists. It’s about engaging domain experts throughout the many stages needed for AI development. It’s about ensuring that our AIs are mimicking the actions of the world’s most expert experts.

But what about at the end of the process, when a team of domain experts, data experts, AI experts and operational experts have all come together to change the world? Are we happy to just let that AI loose on the world without having an independent body check that the AI was developed to the highest standards?

In the case of medical diagnostic software, we currently require new products to undergo a rigorous certification process and AI should be no different. As AIs become more prolific we see a growing number of regulatory frameworks: either voluntary or compulsory, either ethical or technical, either industry- or region-specific. Examples of these include the EU’s guidelines mentioned above, the process of registering Software as a Medical Device (SaMD) and Medical Device Software (MDSW), the Aletheia Framework by Rolls Royce or any of the other emerging frameworks like the NHSX’s AI Buyer’s Guide.

Regulation, to gliff.ai, is not just about bureaucracy, but rather it’s about the documentation and safe-guarding of fair, accountable, transparent and ethical (FATEful) AI. To aid AI product developers, in-house regulatory experts and external auditors, gliff.ai is designing its entire platform so that actions by all parties that interact in the development of an AI are stored in a human-readable auditable log designed to aid companies in proving they meet the requirements of these frameworks.

The missing people in MLOps — let’s get them involved now

So, we’ve identified that there are missing people in MLOps, what can we do? Well, let’s get them involved. And that’s what gliff.ai is all about — completing the MLOps pipeline by bringing domain experts, data experts, AI experts, production experts and regulatory experts together in one environment. Only by doing this can we create amazing AI that really changes the world. If you’re interested in finding out more then drop us a line.