Different from models like ChatGPT, a Chinese research team has developed a new artificial intelligence (AI) system that mimics the workings of brain neurons, paving the way for next-generation, energy-efficient computing and hardware.
Scientists from the Institute of Automation under the Chinese Academy of Sciences (CAS) introduced "SpikingBrain-1.0," a large-scale model fully trained and inferred using China's homegrown GPU computing.
Unlike mainstream generative AI systems that rely on resource-intensive Transformer architectures, where intelligence evolves as the network, computing budget, and dataset grow larger, this new model takes a different approach, allowing intelligence to emerge from spiking neurons.
This model allows for very efficient training on very minimal data volumes.
Using only about 2 percent of the pre-training data required by large mainstream models, the model is able to achieve performance comparable to some open-source models on language understanding and reasoning challenges, the research team said.
By leveraging event-triggered spiking neurons at the inference stage, one variant of SpikingBrain was shown to be able to increase the speed by up to 26.5 times compared to the Transformer architecture when generating the first token from a context consisting of one million tokens.
The ability of this model to handle ultralong sequences provides clear efficiency advantages for a variety of tasks, such as legal or medical document analysis, high-energy particle physics experiments, and DNA sequence modeling.
The research team has open-sourced the SpikingBrain model and launched a public testbed, as well as released a large-scale, bilingual technical report that has been validated by the industry.
"This large-scale model opens a non-Transformer technical path for the development of a new generation of AI," said Xu Bo, director of the Automation Institute. "It has the potential to inspire the design of next-generation neuromorphic chips with lower power consumption."
