Integrating sustainability into the ML lifecycle

Instruction and application

Complete

Sustainability belongs in every lifecycle stage: problem framing, data, training, evaluation, deployment, monitoring and retirement. Regulations and stakeholder expectations increasingly assume you can showwhat you built, with what data, at what cost.

Sustainable design principles (summary)

• Energy-efficient models: pruning, quantisation, distillation; architectures such as MobileNet and EfficientNet when latency and power matter.
• Lean data: deduplication, compression, selective collection, active learning to label the most informative rows.
• Efficient training: transfer learning (for example fine-tuning BERT), AutoML platforms (Google Cloud AutoML, H2O AutoML, AutoKeras), Bayesian or staged hyperparameter search instead of exhaustive grids.
• Scalable architecture: modular pipelines, microservices, elastic scaling with Kubernetes, serverless for bursty inference, versioning to avoid full retrains.

Measuring and monitoring

You cannot improve what you do not measure. Typical indicators:

✨

Energy / carbon:

kWh per training job, estimated CO₂e per 1M inferences, trend lines after optimisations.

✨

Storage and I/O:

deduplication ratio, cache hit rate, bytes moved per experiment.

✨

Model efficiency:

accuracy per parameter, accuracy per millisecond latency, epochs-to-target metric.

✨

Lifecycle view:

periodic assessments of training, serving and cooling impacts (see vendor LCA-style reporting from major clouds).

Key point

Treat sustainability metrics like reliability SLOs: define owners, dashboards and review cadence—otherwise “green AI” slides become shelfware.

Continuous improvement

Sustainability is iterative: models drift, grids get cleaner, hardware improves. Bake in quarterly reviews of model catalogue, experiment hygiene, retirement of unused datasets and refresh of efficiency baselines.

Case study: Foodforecast (demand forecasting vs food waste)

Food waste in bakeries ties to mis-estimated demand. ML forecasting that blends sales history, weather, events and promotions can align production with demand, cutting spoilage and embodied energy from ingredients and ovens. For methodology and lifecycle framing, see the research discussion on ML-based demand forecasting and bakery environmental impacts.

Action item: Quiz

The Foodforecast system demonstrates a creative way AI can reinforce sustainable practices in the bakery industry. Which of the following principles of sustainable AI design is MOST directly addressed by Foodforecast's ability to predict demand accurately?

A) Energy-efficient model design.B) Data minimisation and efficient management.C) Efficient model training and hyperparameter tuning.D) Reducing overproduction and minimising waste through accurate prediction.

An organisation is developing an AI system and wants to integrate sustainable practices from the outset. Considering the principles discussed, which of the following approaches would be MOST effective in ensuring long-term sustainability of their AI system?

A) Focusing solely on achieving the highest possible accuracy during model training, regardless of computational cost.B) Implementing modular and scalable architecture that allows for independent scaling and updates of different components.C) Prioritising the collection and storage of all available data, as it might be useful for future model iterations.D) Measuring only the performance metrics (e.g. accuracy, speed) of the deployed AI model to ensure it meets business requirements.

You are a project manager overseeing the development of an AI-powered customer service chatbot. To apply the principles of sustainable AI development in your role, which of the following actions would you prioritise?

A) Ensuring the chatbot is trained on the largest possible dataset of customer interactions to handle a wide range of queries.B) Selecting the most complex and computationally intensive language model available to ensure the highest level of conversational ability.C) Implementing real-time monitoring of the chatbot's energy consumption and identifying opportunities to optimise its underlying model or infrastructure.D) Focusing primarily on the initial deployment of the chatbot and addressing potential sustainability concerns in later iterations to meet the project timeline.

In progress