Researchers Develop a Framework of AI Training Standards

Scientists in the life sciences, using machine learning in their work, should adopt standards that will allow other researchers to reproduce their results. This is stated in a commentary on the article published in Nature Methods.

The authors proposed three levels of algorithm-accessibility standards: bronze, silver, and gold. Each sets minimum requirements for sharing training materials so that other researchers can trust the work, verify its results, and build on it.

To obtain the Bronze standard, researchers must open access to datasets, source code, and models.

The Silver standard envisages extending the Bronze level with additional information about the system in which the code runs. For other scientists, this information will allow them to reproduce the training process.

To qualify for the Gold standard, researchers must add to their work a “simple one-click button” so that other teams can reproduce the previous analysis with a single command.

According to the authors, the standards are the key to scientific breakthroughs, the expansion of knowledge, and ensuring reproducibility of results from one group of scientists to another.

“Standards will allow other teams to focus on the next breakthrough, rather than waste time reinventing the wheel created by the authors of the initial study,” the paper says.

Casey S. Green, director of the Center for Artificial Intelligence at the University of Colorado School of Medicine, believes that trust in prior work is a significant factor in the research as well.

“All science requires trust — no scientist can reproduce the results of every article they read,” he said.

In July, it was reported that by 2024 in Russia will develop more than 200 standards for AI.

In June, researchers from UNESCO named artificial intelligence and robotics the most popular directions in science over the past five years.

In March, the audit firm KPMG said that the pandemic accelerated the deployment of AI-based systems in various sectors.

Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!