Jakarta, CNN Indonesia —
Researchers at the Chinese Academy of Sciences’ Institute of Automation in Beijing introduced an artificial intelligence system (AI) a new called Spikingbrain 1.0.
This AI is described as a large language model that is ‘like the brain’ of humans. The system is designed using minimal energy and operates on hardware made by China, so it does not depend on the chip made by the United States company, Nvidia.
“Large Language Model (LLM) based on a mainstream transformer faces significant efficiency obstacles: Computing Training Quality with long sequences, and memory of inference that grow linear,” the researchers said in a technical paper that had not been reviewed (non-peer-reviewed).
ADVERTISEMENT
SCROLL TO CONTINUE WITH CONTENT
According to the research team, Spikingbrain 1.0 performs certain tasks up to 100 times faster than some conventional models when trained using data less than 2 percent of what is usually needed.
This project is part of a broader scientific research on neuromorphic computing, the aim is to replicate the efficiency of the human brain that operates only with a power of about 20 watts.
“Our work is inspired by the brain mechanism,” the researchers added.
The core technology behind Spikingbrain 1.0 is known as ‘spiking computation’, a method that mimics the workings of biological neurons in the human brain.
Instead of activating the entire network to process information, as is done chatgpt, most of the spikingbrain 1.0 network can remain quiet. Spikingbrain 1.0 uses an event -based approach where neurons only send signals when triggered specifically by input.
This selective response is the key to reducing energy consumption and accelerating processing time. To demonstrate their concept, the team built and tested two models, which were small with 7 billion parameters and the larger contained 76 billion parameters.
Both are trained using a total of around 150 billion data token, a relatively small amount for this scale model.
The efficiency of this model is seen when handling long data series. In one test quoted in the paper, the smaller model responded to the command consisting of 4 million tokens more than 100 times faster than the standard system.
In different tests, the spikingbrain 1.0 variant shows an increase in speed of 26.5 times compared to conventional transformer architecture when producing the first token from the context of one million tokens.
The researchers reported that their system ran stable for weeks in the configuration of hundreds of Metax chips, a platform developed by Shanghai-based companies, Metax Integrated Circuits Co. This sustainable performance on domestic hardware underlines the potential of the system for application in the real world.
This potential application includes a long and medical legal and medical document analysis, research in high energy physics, and complex tasks such as sorting DNA, all of which involve an understanding of large data sets that really require speed and efficiency.
“These results not only show the feasibility of efficient large model training on non-Nvidia platforms, but also describe new directions for the application and application of models inspired by the brain that are scalable in the future computing system,” the research paper knotted.
(fea)
[Gambas:Video CNN]
