Facebook has developed a system for its artificial intelligence (AI) mechanisms that teach them to forget certain information when it is not relevant to their job performance.. This saves memory and input processing.
The new technology, called Expire-Span, is “the first process of its kind” according to the company, and seeks to make the neural networks of artificial intelligence look like the work of the human brain, which it is constantly forgetting. The data, as reported by the company, the current situation.
“Not all memories are the same: learning to forget with expiration date” (Not all memories are the same: learn to forget with expiration), Is a noun Published work In the past hours. As part of this initiative, Facebook is proposing to incorporate mechanisms so that AI can determine which data is worth keeping and which isn’t, with the goal of improving the system’s operation.
Artificial intelligence mechanisms tend to process information randomly, but when this amount of data is very high, as in the case of long texts or videos, computing costs can become unsustainable.
The technology that the company has created works first by predicting the information most relevant to the functioning of the neural network, and accordingly, it links the data to the expiration date and after that it is erased.
This way, essential information is kept for its operation longer, while irrelevant information is removed quickly, freeing up system memory space to focus on essential tasks.
Distinguishing between important data and what is not stipulated by analyzing past content in context, and using it continuously allows the system to learn from its mistakes and retain more and more important data.
The use of Expire-Span provides benefits for some common tasks of AI mechanisms such as modeling language at the character level, and improves proficiency in long-term context tasks in language, reinforcement learning, object collision and algorithm tasks.
Facebook explained that its technology provides advantages over other systems for saving memory in AI operation such as data compression, Because it does not require “blanking” of important information to operate it and thus does not degrade the results.
In the case of long texts, the system works by memorizing only the most important terms, such as names of people and places, and ignoring other secondary words such as prepositions, which in fact allows for improved AI outcomes.
The platform has confirmed that it will continue to operate so that Expire-Span continues to appear more and more like human memory performance, It will study how different types of memory, such as long-term semantics, are incorporated into neural networks.