ESA_WaveVertical.png

Blog

Increasing the value and consistency of ML models with MLOps

Commentary by Eirik Larsen. Chief Solutions Officer, Earther Science Analytics
17 January 2022

When an underperforming ML model begins to pose challenges for business operations, frustration can quickly set in.

But with MLOps, a practice that seeks to increase the quality of production models, within pre-defined business and regulatory environments, potential issues can be overcome before they begin. Centred around a set of best practices, MLOps is gradually evolving into an independent approach to ML lifecycle management.


With MLOps, the entire lifecycle of a project is covered, from model generation, orchestration and deployment, to health, diagnostic and governance. Ultimately, its usage is seeking to streamline the life cycle of models, delivering multiple benefits across several sectors.


In our latest commentary, Eirik Larsen, co-founder and Chief Solutions Officer of Earth Science Analytics, outlines the positives that will arise from its increased usage, and how our EarthNET software is helping to support   MLOps data-driven decision making.


The end-user must always be considered

At Earth Science Analytics, we have used MLOps for our own development, leveraging the process for deployment in our EarthNET solution. We have seen the first-hand value that this delivers internally to our organisation and to our clients, with rapid model iteration of pre-trained networks also facilitating and supporting us to move quickly into new verticals.

Not only for us, but from healthcare to finance, oil and gas to manufacturing, MLOps practices are helping deliver positives across several metrics.


Those using MLOps want activity to centre around model deployment, using the system to release models with confidence in their outcomes, and with the ability to monitor and retrain – where required – the solutions as they go.


Its usage ensures that the asset integrity of a model is maintained, as the set-up helps to certify model behaviour in-line with regulatory and adversarial standards. Any reproductive challenges are also solved, as users can track, take snapshots and manage their own assets behind the development of a model, enabling increased collaboration, as well as the ability to share ML pipelines. Model packaging and validation also arises from its use, supporting flexible and portable model usage across a variety of platforms.


For me, a separate generated value centres around the ability for a geoscientist and data scientist to combine their expertise and knowledge,   with models updated by feedback and interaction at a much greater speed than was previously possible. This means they can quickly identify what needs to be optimized and what doesn't, ensuring they can focus on specific elements in the dataset while providing an understanding into how the algorithms are being used.


Delivering success with Earth Science Analytics


The delivery of successful internal MLOps requires many steps. There is a requirement for experienced, knowledgeable employees to sit behind each stage for a successful project to be delivered. 


We understood this and wanted to make things easier for our customers, so that’s exactly what we did. While traditional processes would require multiple ML engineers from data scientists, data engineers and software engineers, we allow customers to bypass this completely.


When customers purchase an Earth Science Analytics licence, they do so with the knowledge that they have all available steps of MLOps at their fingertips within functional, easy-to-access software, without the need to recruit their own specialists.

Our customers can use the software in the knowledge that even if they don’t see each step of the MLOps life cycle in operation, they are still using the practices, and that they have the ability to fulfil a full cycle by themselves.


I believe that Earth Science Analytics is making the adoption of MLOps as easy and effective as it can be. For example, our EarthNET solution also   removes what is often referred to as a ‘technical debt.’ This is a concept in software development that reflects the implied cost of additional rework caused by choosing limited solutions for a quick fix, as opposed to taking a better approach that could take longer to develop and deploy.



Connecting users with data


We introduced our EarthNET software to the market to provide a solution that delivers faster, cheaper and more accurate predictions when using AI for predictions of rock and fluid properties in the search for and recovery of oil and gas.


The software helps to connect energy company users with their internal and external data assets with high-performance compute resources and AI-powered geoscience software applications. This connectivity and the integrated applications allow EarthNET users to break out of the data and discipline silos and embrace truly integrated and cross-disciplinary data analytics workflows.



Ensuring models deliver results


The advancements in ML technology are there to be taken advantage of but operating without a lifecycle management framework could result in limited success.


This is why MLOps’ importance should never be overlooked. With MLOps, the entire lifecycle of a project is covered, from model generation to orchestration and deployment, providing the solution to deliver accurate and efficient ML models.


At Earth Science Analytics, we are proud to be one of, if not the first, to provide a platform for MLOps within the oil and gas sector. What’s more, we have the unique position of providing a solution and platform for its use within domain expertise operations. Our goal with MLOps activity is to work with data scientists to develop best-in-class algorithms, with these then being made available for deployment. We have a commitment to putting this technology in the hands of those who need it, helping them to create valuable insights.


We have facilitated users to move at pace and at a reduced cost, ensuring the models and data can get into the hands of those who need it. This is paired with an MLOps approach that is helping to manage, and in many cases even avoid, technical debt.


Our work across multiple sectors is going a long way to close the loop on value creation from models to results, helping to provide real collaborative environments that put data and insights at the fingertips of domain experts and decision-makers.