/Logo%202023%20final%20dunkelgrau.png?width=221&height=97&name=Logo%202023%20final%20dunkelgrau.png "logo")
/HP_Hintergrundbild-01-06-2023_paralax-scrolling-Kreis-dunkel_klein.png)
Whitepaper: Best practices for sustainable Artificial Intelligence in companies
How is environmental sustainability of AI measurable?
What measures are available to developers?
How can responsible managers make the field more sustainable?
Artificial intelligence (AI) is ready for business. In 2021, 35% of companies already reported using AI in their day-to-day business. Another 42% said they were researching AI for themselves. Knowing these numbers, it's no longer possible to assume it's just a trend: artificial intelligence has arrived in companies!
The number of use cases is steadily increasing and the issues to be solved are becoming even more complex. As a result, the size and number of AI models must grow in line with the emerging ideas in companies and the operation of these models must become more and more routine.
Since data-intensive models often have to be trained for days on specialized hardware, this creates an ecological footprint that should not be underestimated and that many are often unaware of.
Already in 2019, model training or model finding in the field of natural language processing generated as much CO2 as five American cars including gasoline consumption during their entire life cycle. And the trend is upward.
The size of the models is continuously increasing. The foundations for sustainable design must therefore be laid sooner rather than later.
Hardware and infrastructure
Ecologically efficient design already starts with the selection of hardware and the design of the infrastructure. Is the effectiveness of energy use in the company's own data center guaranteed? This determines how much energy must be expended in addition to the actual calculations. In addition, the provision of specialized hardware especially for machine learning is an issue.
Model selection
In the modeling process, the developer decides which models are used. Basically, the more complex a model is, the more data and also training time are required. In addition to the restriction to simple models, there are numerous methods from research that lead to more sustainability here. Guidelines for the design of deep learning or the selection of the program library for the implementation of the model are examples of adjusting screws with a significant effect.
Model training and optimization
Elaborate hyperparameter optimizations are performed to optimize the performance of a model on the available data. It is not uncommon for model fine-tuning to take up half of all training runs in model development. An efficient design to find the optimal model candidate faster helps to reduce the footprint in training. Optimization strategies from research and metafeatures support this.
Model operation
The productive use of the models is of great importance in the business context. Here, in addition to (meta) data management and provisioning, predictions and retraining on fresh data also occur. A model optimized in terms of execution can reduce the energy consumption of a prediction by up to 83 %. Intelligent monitoring of input data and model performance also reduces the frequency of subsequent model training.
Organizational framework
Sustainability is always an awareness issue. If the recording of energy consumption or the CO2 equivalent caused is introduced, awareness of the topic increases. With organizational frameworks and standards for the development process, responsibility no longer hangs individually with the developers. Measures are thus implemented uniformly.
Would you like to know more about Machine Learning?
Find interesting articles about this topic in our Blog
Increase Efficiency with Apache Airflow Managed Service Operations
Milestone reached: NextLytics becomes Databricks partner
Databricks and MLflow: A Perfect Match for Scalable Machine Learning
