Discarding irrelevant information to speed up the processing of important information is the goal of a new investigation by the Facebook artificial intelligence team
It may seem negative, but forgetting human abilities is the basis for our brains to function effectively. For example, in this way, we can forget a boring meal in a few hours, but remember the dish that impressed us a few years ago. The ability to discard irrelevant information and remember important content is natural for humans, but it is difficult to achieve through algorithms. This technology has been in demand in multiple computing fields for a long time. Because reducing the data to be processed can improve performance and/or reduce necessary resources. For example, in video game graphics processing, the system usually ignores currently invisible elements to reduce the “weight” of the processing. For example, if one of the characters is behind a wall in a 3D game scene, the graphic information of the character can be discarded and processed only when it is visible. A technology that can make the game smoother and reduce the “weight” of the work to be done by the graphics processor.
According to information disclosed by Facebook, researchers in the company’s artificial intelligence field are developing a new method of deep learning with the ability to “large-scale forgetting”. This algorithm is called Expire-Span and works “based on predictions of the most relevant information for the task.” More specifically, Expire-Span assigns expiration dates to data based on the relevance of the data. After this date, messages will gradually begin to be deleted. The result is a significant reduction in the data to be processed, thereby simplifying the processing of artificial intelligence algorithms. This is very important in an era when AI systems must process large amounts of information.
The attempt to imitate human memory is obvious, but the researchers themselves have caught people’s attention: “Although Expire-Span focuses on the memory of past experiences, there are many other types of human memory. Semantic memory, for example, is used to store general and Factual information. As the next step in our research on AI systems that are more like humans, we are studying how to integrate different types of memory into neural networks. Therefore, in the long run, we can make AI closer to human memory, Its learning resources are much faster than today’s systems. We believe that Expire-Span is an important step towards these future AI innovations.
Ø Facebook blog The new technology is introduced in detail.