AI that self-evolves faster than human coding?

AutoML-Zero is an artificial intelligence developed by Google’s engineering team to automatically discover the optimal machine learning algorithm using only basic mathematical operations.

For the past few years, Google has been focusing on developing a system that makes it easy for people without knowledge to use machine learning. However, there was a problem that human knowledge was required to adjust the algorithm.

The newly announced AutoML-Zero by the Google Engineering Team led by Stanford University’s Quoc V.Le, an artificial intelligence expert, is a basic mathematical operation that starts from the absence of any programs or programs to the level of high school students. It is an AI that makes a program that processes machine learning tasks with only one.

AutoML-Zero constructs the underlying algorithm using loose approximations and randomly combines mathematical operations to generate 100 refined candidate algorithms. Afterwards, simple tasks such as discriminating between truck photos and dog photos are used to hide the best performance among the candidate algorithms for the improved version. It also makes a mutant version that replaces and edits some of the best algorithm code so as not to hinder evolution. The algorithm created in this way stores the algorithm with good performance, including the mutant version, but deletes the others. It is said that the series of processes borrowed Darwin’s theory of evolution, including survival of the fittest.

AutoML-Zero writes and verifies tens of thousands of algorithms per second until an optimal algorithm is found. The paper suggests the possibility that AutoML-Zero can go beyond classical machine learning algorithms such as neural networks. The research team believes that by increasing the types of mathematical operations, it is possible to create algorithms that human programmers never thought of by consuming computer resources. Although only simple AI systems can still be created, the research team is trying to improve, such as including human-discovered code, rather than creating an algorithm from scratch, saying that it will be possible to quickly create complex algorithms. Related information can be found here .