Researchers from Facebook have presented a new method of storing information by artificial intelligence (AI). Each piece of data is automatically assigned an expiration date based on its importance.
The researchers explained that the models effectively remember any information. This is usually good, but in some cases it is harmful, because AI does not separate important and secondary data and processes them with the same efficiency, spending about the same amount of data. But as the amount of information available increases, systems are forced to spend more and more of their limited computing resources to process it. Facebook researchers hope to help future AIs be more mindful by giving data an expiration date.
This technique is called Expire-Span and will help neural networks efficiently sort and store information that is most suitable for performing their tasks. Expire-Span works by first predicting what information will be most useful to the network in a given context, and then assigning an expiration date to it. The more important, according to Expire-Span, this or that information, the longer the shelf life will be for the data. Neural networks will be able to keep relevant information for a longer time, constantly freeing up memory, forgetting irrelevant data.
“Expire-Span calculates an expiration date for all types of data and does so every time a new piece of information is presented to it. This tag determines how long information is retained in memory, the researchers explained. “The gradual decay of a piece of data is the key to preserving important information without blurring it.”
Although the research is still in its early stages, scientists are confident that not many steps are needed to improve the neural network. In the future, the team hopes to develop an even closer human memory, but capable of assimilating new information much faster than current technology allows.