Logo Zéphyrnet

DeepMind sets sights on automating the discovery of new algorithms

Date :

DeepMind, propriété de Google, a appliqué des techniques d'apprentissage renforcées à la multiplication de matrices mathématiques, battant certains algorithmes créés par l'homme qui ont duré 50 ans et travaillant à l'amélioration de l'informatique.

Fondé à Londres en 2010, DeepMind est devenu célèbre pour avoir battu le champion du monde au jeu de société Go avec son AlphaGo l'IA et relever le défi époustouflant du repliement des protéines avec AlphaFold.

Dans un mouvement de roues dans les roues, il a depuis jeté son dévolu sur les problèmes mathématiques eux-mêmes.

A successful project has been to automatiser la découverte of algorithms which act as shortcuts when multiplying matrices – the cause of headaches for many a teenage math student.

For years, mathematicians have been applying algorithms to these complex array multiplications, some of which are used in computer science.

DeepMind researcher Alhussein Fawzi and his colleagues used deep reinforcement to rediscover earlier algorithms and find new ones. The technique created a system, dubbed AlphaTensor, which plays a game in which the goal is to find the best approach to multiplying two matrices. If the AI agent does well, it is reinforced to make future success more likely.

In this case, the agent takes on puzzles in the form of a 3D tensor or a grid of numbers, which it must complete in the fewest moves. Each step represents a move in solving the matrix-based puzzle, which might contain trillions of possible moves.

Fawzi told a press briefing this week that mapping out the space of algorithmic discovery was tough work, although navigating it was even more difficult. Nonetheless, the resulting research developed new algorithms for problems which have not been improved on in more than 50 years of human research, he said.

The researchers claim the technique could benefit computational tasks that use multiplication algorithms as well as demonstrate how reinforcement learning can be used to find new and unexpected solutions to known problems, while also noting some limitations. For example, predefined components are necessary to avoid the system missing a subset of efficient algorithms.

Les sceptiques peuvent pointer vers l'application d'AlphaFold, qui a promis des percées dans la découverte de médicaments via la recherche sur les protéines soutenue par l'IA. Bien que le modèle ait prédit presque toutes les structures protéiques connues découvertes, son capacité à aider scientists discover new drugs remains unproven. ®

spot_img

Dernières informations

spot_img