Google engineer announces AI 'AutoML-Zero' that self-evolves faster than humans code
Google's engineer team announced artificial intelligence (AI) ' AutoML-Zero ' that automatically discovers the optimal machine learning algorithm using only basic mathematical operations.
[2003.03384] AutoML-Zero: Evolving Machine Learning Algorithms From Scratch
Artificial intelligence is evolving all by itself | Science | AAAS
https://www.sciencemag.org/news/2020/04/artificial-intelligence-evolving-all-itself
Google Engineers'Mutate 'AI to Make It Evolve Systems Faster Than We Can Code Them
https://www.sciencealert.com/coders-mutate-ai-systems-to-make-them-evolve-faster-than-we-can-program-them
Google has been focusing on developing a system that allows people who do not have knowledge of ' AutoML ' to easily use machine learning for many years, but human knowledge is required to adjust some algorithms I was left with the problem.
Newly announced by Google's team of engineers led by machine learning and artificial intelligence expert Stanford University Quoc V. Le, 'AutoML-Zero' starts from a random program or no program at all. , AI that creates programs to process machine learning tasks with only basic mathematical operations that are learned by high school students.
AutoML-Zero constructs a base algorithm using loose approximation and randomly combines mathematical operations to generate 100 'improved candidate' algorithms. Then use simple tasks like distinguishing truck photos from dog photos to identify the best performing algorithms among the improved candidates. It also creates a 'mutated' version that replaces, edits, and deletes some of the code of the best algorithms to prevent the dead end of evolution. As for the algorithms created in this way, good algorithms, including mutant versions, are saved, but other algorithms are deleted. A series of cycles is described by Science as 'borrowing Darwin's Theory of Evolution , including Harvard Spencer's The Survival of the Fittest '.
AutoML-Zero will create and test tens of thousands of algorithms per second until the optimal algorithm is found. Published papers suggest that AutoML-Zero may produce classical machine learning algorithms such as neural networks .
The research team believes that by increasing the types of mathematical operations and spending computer resources, it is possible to create algorithms that human programmers have never thought of. At the time of announcement, only a simple AI system can be created, but the research team said, `` We will be able to quickly create complex algorithms '', `` Instead of generating algorithms from scratch, we will incorporate human-found code. We promise that we will respond in the future to such a point.
AutoML-Zero is published on Google-Research on Github.
google-research / automl_zero at master · google-research / google-research · GitHub
https://github.com/google-research/google-research/tree/master/automl_zero
Related Posts:
