In their latest research, the team behind Google’s DeepMind has revealed an artificial intelligence platform capable of tackling real-world programming challenges.
The new AI system can participate in typical computer programming competitions at a similar level to humans. It is capable of producing software code aimed at solving different tasks. Apparently, its reasoning is based on similar principles that are being used by programmers. And, this artificial intelligence works quite well – it can achieve scores in the top 54.3% of participants.
The team presented their work in the latest edition of the Science journal.
How does programming with AI work?
Previously, the DeepMind team has presented an AI system that is able to create better versions of existing mathematical algorithms. Their latest work is a step further.
The code-generating AI is built using a transformed-based neural network model. The researchers note that this type of machine learning model is particularly suitable for solving similar programming tasks. However, earlier attempts were less successful due to issues related to the complexity of problems at this level of coding, especially when longer software code was necessary.
The new platform works well. Interestingly, it performs its functions as intended without prior structural knowledge of algorithms, or programming languages. How it accomplishes such an incredible feat?
Researchers explain in their article, that the new program uses a generative approach, when it produces millions of diverse versions of the code. This generation is done using specially trained transformer-based neural networks. Later, the best-performing results are automatically selected.
This is the first time when artificial intelligence performed relatively well in a programming-oriented competition.
(Except for the headline, this story has not been edited by PostX Digital and is published from a syndicated feed.)