Sustainable AI: Reusing Models For Sustainability

[Mattab]

Re-Usable Models

We’ve reached the end of our Sustainable AI series, with one final aspect to look at before we take all that we’ve learned and start to create truly sustainable AI models. In this series we have:

• defined what we mean when we talk about sustainable AI,
• compared software engineering against data science,
• looked at repeatability and how it can add efficiency to the process,
• explained what actually happens to the data involved in the process and how to avoid data duplication problems,
• explored how to effectively tune models,
• advised ways to avoid technical debt by staying on top of our data.

All these blogs can be found here, and form a set of useful stepping stones to ensuring AI models aren’t just created in a sustainable way, with a low carbon footprint, but are sustainable in their operation, with longevity built in.

There is one question that we should be asking before we begin to build, however – “Do I need to create a new AI model?” Building an AI model from scratch every time you need to solve a problem is time consuming, complex and expensive. Additionally, it goes against our theme of this series, resulting in unnecessary carbon emissions from the development, build and tuning process. There are countless problems that have already been tackled by AI and made available for use. With a multitude of models out there already, can we minimise our impact by using one that already exists? Reusing a model is a key step in ensuring longevity of an AI system, avoiding the need to go back to the start every time we identify an opportunity to solve a problem with AI. Model reuse and pretrained models are great ways of tackling complex tasks, while avoiding some sustainability challenges.

Research first

There is a broad ecosystem of prebuilt models which is simple to use and easy to access. You may just need to do your research first to ensure the chosen model is best suited to your needs. There are many portals to access services, software and support for AI projects, providing organisations who are looking to build and deploy solutions with the tools they might need to develop these solutions more quickly.

As an example, NVIDIA GPU Cloud or Hugging Face which are catalogues that hosting models and datasets.. NVIDIA AI Foundations are a set of customisable cloud services for generative AI projects. While Hugging Face focuses on providing a community focused at building, training and deploying models based on open source projects code and techniques. These foundations form a set of customisable services for generative AI projects, and cover a range of areas, such as::

• Text generation, using NeMo from NGC or Mistral-7B from the open-source community; offering a fast way to build, customise and deploy generative AI. These approaches are easy, and both cost and carbon effective, this is a great service to explore.
• Visual content using LLaVA from the open-source community or Picasso from NGC to provide a model architecture to build, customise and deploy multimodal implementations for text-to-image. 
• Generative AI is also permeating into disciplines such as biotech with BioNeMo from NVIDIA or BioMistral from the open-source community becoming means of simplifying and accelerating the training of models using organisational data and scaling the deployment of models for drug discovery applications.

There is a broad collection of open-source models that can be used as the foundation for your AI activities. Each of these approaches is adaptable and can be fully tailored to your specific needs.

Creating an AI strategy

A well-constructed and fully thought-out strategy will include looking how and when you plan to use your AI models. This will ensure you create them with longevity in mind, meaning you can reuse them efficiently and gain maximum value whilst keeping their carbon footprint as low as possible. A robust plan will ensure sustained benefits, and if I refer back to my very first blog in the series where I referenced the ancient Chinese military book, ‘The Art of War’¹, “Tactics without strategy is the noise before defeat. Strategy without tactics is the slowest route to victory”. It’s essential to have both, in the right order, at the right time.

This strategy should also ensure you have places to store the AIs and features you create, building a culture of reuse. Sharing best practice and encouraging this circular thinking is also a great way of organically building in sustainability to your approach.

AI is already a well-travelled road when it comes to identifying and solving problems. Working with trusted partners can be a great benefit as they will have that historic knowledge and may have already identified certain pitfalls (which you can uncover in our AI Pitfalls series). They will also be able to guide you to comprehensive ecosystems that will allow you to see if your specific problem has already been tackled, reducing duplication in a way that is both beneficial to you and the planet.

¹The Art of War, Sun Tzu, 2010, Capstone Publishing.

Matt Armstrong-Barnes
Hewlett Packard Enterprise

twitter.com/HPE_UKI
linkedin.com/company/hewlett-packard-enterprise
hpe.com/uk