Google DeepMind announces ``FunSearch'', an AI system that uses large-scale language models to output new solutions to mathematical science problems more skillfully than mathematicians



Google DeepMind , Google's AI development division, announces FunSearch , an AI system that uses large-scale language models (LLMs) to solve unresolved problems in mathematical sciences and generate new knowledge. Did. Google DeepMind CEO Demis Hassabis said, ``We are excited about FunSearch's breakthrough.''

FunSearch: Making new discoveries in mathematical sciences using Large Language Models - Google DeepMind
https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/



DeepMind AI outdoes human mathematicians on unsolved problem
https://www.nature.com/articles/d41586-023-04043-w

Mathematical discoveries from program search with large language models | Nature
https://www.nature.com/articles/s41586-023-06924-6

DeepMind AI with built-in fact-checker makes mathematical discoveries | New Scientist
https://www.newscientist.com/article/2407897-deepmind-ai-with-built-in-fact-checker-makes-mathematical-discoveries/

AI scientists make 'exciting' discovery using chatbots to solve maths problems | Science | The Guardian
https://www.theguardian.com/science/2023/dec/14/ai-scientists-make-exciting-discovery-using-chatbots-to-solve-maths-problems

LLM, the technology underlying chat AI such as ChatGPT and Bard, is a powerful neural network that learns language patterns, including computer code, from vast amounts of text and other data. Since ChatGPT was introduced in 2022, various AI tools using LLM have appeared.

'FunSearch' was developed by Google DeepMind using such LLM. The name FunSearch stands for 'searching in the function space' and was built as an AI tool that can create solutions to problems in the form of computer programs.

FunSearch is combined with an 'evaluator' that automatically ranks programs by how well they perform, and FunSearch is said to be able to evolve poor programs into more powerful programs that can discover new knowledge. It has been.

'I'm excited about the latest breakthrough announced today in Nature,' Hassabis said on his X account. 'It's the first system that can perform testable new discoveries. We're using it to find problems more efficiently. It's a solution to the classic bin-packing problem .'




The Google DeepMind research team entered two puzzles into FunSearch. One is `` Cap set '', which has been considered an unsolvable mathematical problem for many years, and FunSearch has succeeded in outputting multiple excellent solutions that exceed the answers devised by mathematicians.

Another math problem entered into FunSearch is the 'bin packing problem,' which seeks to find the best way to pack items of various sizes into bins. FunSearch has also succeeded in discovering a better approach to the bin packing problem than the algorithms devised by humans to date.

Tim Gowers, a professor of mathematics at the University of Cambridge who was not involved in the study, said: ``Over the past few years, there have been several interesting examples of human mathematicians collaborating with AI to solve unsolved problems.'' This work has the potential to offer a different and very interesting approach to collaboration between researchers and AI, allowing mathematicians to efficiently discover clever and unexpected structures. Even better, the algorithms FunSearch outputs are human-interpretable.'



The Google DeepMind research team is investigating the range of scientific problems that FunSearch can handle. The main limiting factor may have a more direct impact on computer programmers. Over the past 50 years, coding has improved significantly as humans create more specialized algorithms. Pushmeet Kohli, Head of AI for Scientists at Google DeepMind, said: 'FunSearch will revolutionize the way people approach computer science and algorithmic discovery. It will definitely help us push the limits of things,' he says, and is looking forward to FunSearch's future success.

'FunSearch doesn't generate solutions; it generates programs that find solutions. A solution may not provide any insight into how to solve other related problems. However, the 'solution finding program' that FunSearch outputs is human readable and interpretable. , and hopefully that generates ideas for the next problem, and the one after that,” he says of FunSearch.

Related Posts:

in Software, Posted by logu_ii