Resource-Efficient AI for Sustainable Deployment

Objective

The project aims to address the increasing energy consumption and carbon footprint of large-scale AI/NLP models by exploring resource-efficient methods. The goal is to reduce the computational resources without compromising performance.

Content

The project focuses on developing and testing methods for resource-efficient training and deployment of large language models. By combining more data-efficient methods for training with model compression for deployment, we aim to reduce resource requirements over the entire life-cycle of an AI/NLP model. In addition, we will create a novel conceptual framework for life-cycle analysis of resource efficiency.

Methodology

For data-efficient training we will explore a generative-discriminative approach that can be made efficient thanks to the technique of non-residual attention developed by researchers at RISE. For resource-efficient deployment we will investigate task-agnostic and task-specific distillation methods. For the evaluation of resource-efficiency, we will measure running time, power consumption, and number of floating point operations and correlate these measurements with different product measures, including product-as-performance and product-as-output.

Impact

The project is expected to deliver the following results:

• More efficient training and deployment of AI models, especially large language models.
• More adequate metrics of resource-efficiency for AI models, especially large language models.
• Better support for sustainable and efficient AI implementations for companies and other organizations.

In this project, we aim to take steps towards reducing energy consumption and optimizing resource utilization in the AI field. We want to pave the way for a more sustainable and resource-efficient future for AI and support companies and organizations in their transition to AI-driven solutions.