Facebook has taught AI to treat waste like spoiled milk
Computers are perfect for remembering everything we teach them. Usually, that’s fine; you do not want the systems that keep your medical or financial records to start losing 1s and 0s (Well, maybe those who follow your credit card debt, but otherwise). However, these systems do not discriminate between the two, meaning that everything that is changed is by the same power. But as the amount of information grows, AI systems need to use less complex tools to solve them. Researchers on Facebook are hoping to help future AIs manage better .
It is called Expire-Span and is designed to help networks to better organize and store information that is most relevant to the tasks assigned to them. Expire-Span works by first predicting which information would be most useful on the network in the given time and giving the last date to the passage. The essence of Expire-Span thinks that the knowledge is that, far from giving the expiration date, Angela Fan and Sainbayar Sukhbaatar, research scientists at FAIR, explained in . As a result, networks are able to store important information over a long period of time while eliminating memory space by “forgetting” unnecessary information. Each time a new information is added, the system re-evaluates not only its relevance but also the relevance of existing information. This will also enable AI to learn to use memory that is more accessible, which makes things better.
Forgetting, for at least AIs, can be difficult because doing so is . Like the 1s and 0s that make up the AI number, the system is able to recall information or not. As a result, the optimization of such a sophisticated system becomes more difficult. Early attempts to address the problem involved suppression of non-essentials in order to take up less space in memory but the attempts were not made because the suppression brought about the “immediate appearance” of the message, according to Fan and Sukhbaatar.
“Expire-Span calculates the maximum amount of information in each secret state, each time new information is provided, and indicates that the information is stored as a memory,” he explained. “The temporary loss of knowledge helps you to keep important information in check. And the learning machine allows the type to adjust the span size if needed. Expire-Span calculates predictions based on information learned from multiple sources and based on memories. ”
Although we are still at the beginning of this study, “As the next step in our research using human AI techniques, we are learning how to create different memories in these networks,” the research team wrote. In the future, the team hopes to expand its approach to human memory but it is possible to learn new things faster than current technology allows.
All the products that Engadget selected were selected by our management team, independent of our parent company. Some of our articles include helpful links. If you purchase one of these links, we will be able to make a donation.