Facebook has developed a technology to enable artificial intelligence to be able to forget little necessary information, just like the human brain does. The technology will help improve the functioning of neural networks, saving on costs of memory Y data processing.
Named ‘Expire-Span’ (‘Expiration period’), this deep learning method, the “first of its kind”, provides a memory capacity “by an order of magnitude greater” than that of standard models of neural network functioning, as evaluated by Facebook in a release.
The system predicts which information is most relevant and which least for a given task, based on which it assigns a expiration date to each unit. With the arrival of that date, the associated information is deleted. Dispensing with less relevant information makes it faster, by free up memory to facilitate data processing on a larger scale.
The Expire-Span determines the relevance of the information by analyzing the contextual background data and surrounding memories, while error learning refines the determination of the fundamental data.
Expanded to tens of thousands of information units, the system is able to retain fewer than 1,000 of them and thus achieve more efficient performance than alternative methods. In addition, it is distinguished by distancing itself from the common practices of compress data as they age, which although it expands the scope as far as the past is concerned, it only provides blurry versions of those memories.
The new technology offers advances in some common tasks of using artificial intelligence, such as language modeling in terms of characters, and improves efficiency in long-term context tasks in terms of language, reinforced learning, object collision and algorithmic tasks .
Inspired by the human way of remembering, providing easy access to pertinent memories and avoiding flood with details of little value, Facebook assures that it will continue working to make its technology more and more similar to human memory and incorporate other of its various mechanisms into neural networks.
If you liked it, share it with your friends!