Project SHARP

Summary

SHARP is a project created by a consortium of schools and research institutes (Inria, CNRS, CEA, ENS Lyon, etc.) as part of France 2030.

The objective of SHARP is to propose a theoretical and algorithmic framework for leveraging existing knowledge and modern concepts of predictor and/or algorithm sparsity. Among other things, SHARP will explore two concrete applications:

1. frugal training of compact transformers

2. representation learning on small, unlabeled datasets

Thus, SHARP aims to concentrate research on architectures, learning principles, and data to define the most frugal learning methods while maintaining model performance.

The major challenge of the SHARP project is to make a significant leap in frugality by designing, analyzing, and deploying intrinsically efficient models (whether neural or not) capable of achieving the versatility and performance of the best models while requiring only a fraction of the currently needed resources.

To achieve this level of AI frugality, SHARP will rely on three pillars: frugal architectures, frugal learning principles, and learning with small and rare datasets.

The editorial perspective

The SHARP project was launched in March 2024. Scientific publications from SHARP are already available online on the project’s website.

It will be highly relevant to follow SHARP’s progress and announcements over time. The project has a clear ambition:

- To develop highly technical frugal AI models

- To make these solutions deployable in real-world applications

The level of technical expertise in SHARP’s frugal AI research is remarkable, and it is expected to produce innovative solutions in this field.