Building AI for the future
Artificial intelligence (AI) has transformed industries and societies, offering unprecedented capabilities and efficiencies. However, as the use of AI proliferates, questions around how to create a robust and sustainable strategy for AI have come to the forefront, writes Davin Cody, Chief Technology Officer, HPE Ireland.
Especially for those who are still at the beginning of their AI journey, it is difficult to get it right from the start. To get a better idea of how we can future-proof AI projects, we need to understand what aspects of the lifecycle of an AI project, model or platform are pivotal to their long-term success.
Drawing from years of providing AI-native infrastructure, designing energy-efficient IT solutions, engaging with and supporting open-source projects, and being part of the AI Infrastructure Alliance, we at HPE learned that to ensure the viability of an AI project beyond momentary success, we need to look at four key aspects across the whole lifecycle: the environmental impact, longevity, agility, and reusability.
Environmental impact
The first thing that comes to mind when we talk about sustainability is the actual environmental impact, and unfortunately the carbon footprint of AI systems is not small. An AI project needs a clear strategy to not just ensure long-lasting benefits, but also maximum efficiency, including objectives like energy efficiency to mitigate environmental impact. This includes aspects related to the design of AI systems, covering technology infrastructure, through to monitoring and controlling. These decisions need to be aligned with the overall sustainability targets of the organisation to ensure that they can fulfil their commitments.
As you start to outline your project, make sure that the project team consist of representatives from stakeholder groups across the organisation as well as a spectrum of skill sets, including IT, data, and software engineering experts. The reason for the latter is that software engineering in particular aims at building efficient, sustainable, maintainable, and supportable applications. With operations being performed hundreds of millions of times (or more), for example during the AI model training and, a small performance hit has a big impact. The adoption of software engineering practices into your data science projects helps to avoid these challenges.
The efficiency advantages will become apparent with the implementation of machine learning operations (MLOps). MLOps can help save time, resources and reduce carbon output by minimising the number of steps that must be carried out manually, therefore making it repeatable through automation. As with all software engineering, AI has a defined lifecycle, comprising a collection of steps that moves an idea through to production deployment. AI breaks down into five key activities within this lifecycle: data processing, pipelining, model development and optimisation, model deployment, and monitoring. Having a clear strategy on these will enable effective automation for repeatability, and saves time and resources, leading to a successful and more efficient AI project.
As mentioned above, without data there is no AI. This means it is essential to plan right from the start how data is going to be handled, deduplicated, and metadata stored. It is common for organisations to hoard data, gathering it all up because you do not know yet what is valuable. However, to stay efficient, it is important to have a clear data selection criteria and disposal processes, and core information on the data – where it came from, how it was transformed and by whom, why it was transformed, how long it was kept for and how it will be disposed of. This process should be part of your wider existing data strategy, to ensure your data is handled consistently and efficiently. Truly understanding your data will ensure that your model is robust and sustainable, also benefiting the longevity of your project.
System longevity
With the goal being to build AI that offers continuous benefits, it is vital to create it with long-term aims in mind that align to your AI strategy. If designed for the short-term use or goals only, you will soon need to train it again or build a new one, increasing costs and your overall carbon footprint, with model tuning being the most resource intensive part of an AI lifecycle. However, if designed to tackle real and long-term business challenges, you will derive more value from the projects with a lower impact.
Agility
One challenge you will encounter when building your AI system while aiming for agility is data debt, which can accumulate rapidly. In essence, this means if you do not keep on top of its classification, any problem gets bigger exponentially. Therefore, it is essential that you follow a process, to keep your AI systems in up to date.
Model reusability
Model reuse and pretrained models are great ways of tackling complex challenges. They allow you to avoid additional AI tuning, which, as we already touched on above, can have serious impact on cost and carbon footprint. It is essential to have a detailed plan to ensure you get the most out of your models. At the same time, it is important to keep in mind that the conceptualising an AI model for reusability increases development costs in the short term. Especially when building interdependent AI systems, the management of these systems can turn into a time- and resource consuming challenge. Therefore, building AI with a firm plan and purpose is key.
To summarise, it is crucial, when building AI systems, that we architect them with the future in mind. By prioritising future-oriented and efficient AI practices, we can contribute to a more sustainable planet while harnessing the transformative power of AI. As we navigate the intersection of technology and environmental consciousness, the pursuit of environmentally sustainable AI represents a commitment to building a future where innovation coexists harmoniously with the well-being of our planet.
To find out more about how you can build sustainable AI, check out our video series at https://www.hpe.com/uk/en/solutions/artificial-intelligence/nvidia-collaboration.html#SustainableAIEpisodes.
