Technical avenues for more frugal AI

Summary

A list of potential solutions for frugal AI. This resource provides a concise overview of various solutions discussed in other referenced materials.

It is not intended to provide an expert analysis, but rather to highlight possible approaches to reduce AI’s environmental impact:

1 - Purpose of the service

2 - Edge computing

3 - Local processing

4 - Synthetic data

5 - Transfer learning

6 - Partial retraining of AI models

7 - Limiting the dataset size

8 - Reducing the number of model parameters

9 - Minimizing AI memory consumption using neuromorphic chips, for example

10 - Algorithm compression

11 - Pruning: a technique that removes certain neuron connections in a neural network

12 - Distillation: transferring knowledge from a larger model to a smaller one while maintaining most of its accuracy

This initial list is intended to be expanded and updated over time.