Google DeepMind has discovered a way to create AI algorithms that run faster than those coded by humans, which could lead to more cost-effective software development and computing that is more efficient and sustainable, according to the Alphabet company. The breakthrough, detailed in the journal Nature, is called AlphaDev. It uses a form of machine learning called reinforcement that allows computers to build on their successes, honing strategies independent of human programmers. In this case, faster algorithms were developed for computer-science functions like sorting and hashing.
Sorting algorithms are used in functions ranging from the institutional management of financial data to the filtering that results in search engine rankings. Hashing algorithms convert data of any size into a unique string of fixed-size, making them useful for tasks like identification, indexing and security. Because these techniques are widely used, faster algorithms could significantly reduce cost.
“Fundamental algorithms such as sorting or hashing are used trillions of times on any given day,” according to the Nature abstract on the AlphaDev paper. “As demand for computation grows, it has become critical for these algorithms to be as performant as possible.”
AlphaDev “created algorithms that, when translated into the standard programming language C++, can sort data up to three times as fast as human-generated versions,” the Nature article explains.
“We were a bit shocked,” said DeepMind computer scientist Daniel Mankowitz, who led the research “We didn’t believe it at first.”
When it comes to data sorting, milliseconds are crucial — as anyone who’s waited for search results, or a web page to load, can attest. “AlphaDev was up to 70 percent faster for smaller sorting tasks and 1.7 percent faster for large-scale sorting tasks,” according to The Wall Street Journal, which says for hashing functions, “AlphaDev discovered an algorithm that was 30 percent faster in the 9 to 16 bytes range.”
AlphaDev is based on the technology that powers DeepMind’s AlphaZero, an iteration of AlphaGo — famous for mastering the complex Chinese board game Go.
Google has made AlphaDev’s sorting and hashing algorithms available in open-source libraries. It has also used the technology in-house to optimize network resources across Google and Alphabet as well as to make UK-based DeepMind’s computer systems more efficient.
AlphaDev Discovers Faster Sorting Algorithms, Google DeepMind, 6/7/23
Google DeepMind’s Game-Playing AI Just Found Another Way to Make Code Faster, MIT Technology Review, 6/7/23
Deepmind Discovers Sorting Algorithms That Can Revolutionize Computing Foundations, VentureBeat, 6/7/23