AI is evolving- Actually. Researchers have developed software that borrows concepts from Darwinian evolution, including “survival of the most powerful,” to build AI systems that develop for generations without being installed by anyone. The system is a replica of decades of AI research in a matter of days, and its designers think that one day, it can find new ways for AI.

Ricko Miikkulainen, a computer scientist at the University of Texas, Austin, who was not involved in the project, says: “While most people are taking baby steps, they are taking a big step towards the unknown.

Creating an AI algorithm takes time. Take neural networks, the most common type of machine learning used to translate languages and drive cars. These networks mimic the formation of the brain and learn from training by altering the ability to communicate between synthetic neurons. The lower limbs of neurons perform specific tasks – for example detecting traffic signals – and researchers can spend months working on how they connect to work together.

In recent years, scientists have completed the process by changing some of the steps. But these systems still rely on cutting together man-made circuits. That means that the output is still limited by developer ideas and their existing guesses.

So Quoc Le, a computer scientist at Google, and his colleagues have developed a program called AutoML-Zero that is able to develop AI systems with human input, using only the basic mathematical concepts a high school student will know.

The system detects algorithms using open evolutionary equations. It starts by building a group of 100 algorithms by randomly combining mathematical operations. It then tests them for a simple task, such as the problem of seeing a picture where they have to decide whether the image shows a cat or a truck.

Related: UV Light-A New Source of Green Fuel

In each cycle, the system compares the algorithms’ performance with the hand-built algorithms. Copies of top players are “modified” by replacing, editing, or deleting some of their code to create a small difference for the best algorithms. These “children” are being added to group, while older programs are becoming more popular. The cycle repeats.

The program builds thousands of these sites at once, allowing you to work with tens of thousands of algorithms per second until it finds the best solution. The system also uses tactics to speed up the search, such as periodically exchanging algorithms between people to prevent any end changes, and automatically removes duplicate algorithms.

In the first paper published last month on arXiv, researchers suggest that this approach could stumble across many machine learning algorithms, including neural networks. The solutions are simple compared to today’s highly developed algorithms, admits Le, but say the work is policy-proof and hopes it can be allowed to create more sophisticated AI.

Also, Joaquin Vanschoren, a computer scientist at Eindhoven University of Technology, thinks it will be some time before this approach is consulted with the state. One thing that could improve the system, he says, is not asking for it to start at the beginning, but instead investing in other strategies and strategies that people have discovered.

That’s what Le plans to work on. Focusing on smaller problems instead of algorithms also holds promise, he adds. His team published another paper on arXiv on April 6 that used the same method of reconstructing a popular object created in multiple neural networks.



Please enter your comment!
Please enter your name here