DeepMind disclosed the FunSearch training method, which allows AI models to perform discrete mathematical calculations

Google DeepMind announced a model training method called “FunSearch” on December 15. It is said that the model can solve a series of "complex problems involving mathematics and computer science", including "upper-level problems" and "bin-packing problems"

The content that needs to be rewritten is: ▲ Picture source Google DeepMind (the same below)

FunSearch model training method is reported to have introduced a system called "evaluator", which is used to judge the creative problem-solving methods output by the AI ​​model. Through repeated iterations, this method can train an AI model with stronger mathematical capabilities

Google DeepMind used the PaLM 2 model for testing. The researchers established a dedicated code pool, used a series of questions as input to the model in the form of code, and set up an evaluation process. In each iteration, the model automatically selects problems from the code pool, generates creative new solutions, and submits them to the evaluator for evaluation. The best solution will rejoin the code pool and start the next iteration

FunSearch training method was mentioned in the IT House report that this method performs well in "discrete mathematics (Combinatorics)", and the trained model can easily solve extreme value combinatorial mathematics problems. In a press release, the researchers describe a procedural approach to the model's calculation of "the upper-level problem, a central problem in mathematics involving counting and permutations."

In addition, researchers also successfully used the FunSearch training method to solve the "Bin Packing Problem", which is a problem of "putting items of different sizes into the minimum number of containers". FunSearch is the "Bin Packing Problem" Provides an "on-the-fly" solution, generating a program that "automatically adjusts based on the existing volume of the item."

Researchers mentioned that compared with other AI training methods that use neural networks to learn, the output code of the model trained by the FunSearch training method is easier to check and deploy, which means it is easier to be integrated into the actual industrial environment. middle.

The above is the detailed content of DeepMind disclosed the FunSearch training method, which allows AI models to perform discrete mathematical calculations. For more information, please follow other related articles on the PHP Chinese website!