For the first time, a team of Chinese scientists has confirmed that multi-modal large language models based on artificial intelligence technology can spontaneously form a system of object concept representation highly similar to that of humans. In other words, AI can spontaneously form human-level cognition.
According to a report by China News Service, this research was jointly completed by the Neural Computing and Brain-Machine Interaction Team of the Institute of Automation, Chinese Academy of Sciences, and the Center for Excellence in Brain Science and Intelligence Technology of the Chinese Academy of Sciences. The related paper was published online on the 9th in the international professional academic journal "Nature Machine Intelligence."
It is introduced that humans have the ability to conceptualize objects in nature, and this cognitive ability has long been considered the core of human intelligence. When people see "dog," "car," or "apple," they can not only recognize their physical characteristics (size, color, shape, etc.) but also understand their functions, emotional value, and cultural significance. This multidimensional concept representation forms the cornerstone of human cognition.
In recent years, with the development of large language models such as ChatGPT, the fundamental question of whether these models can develop human-like object concept representation from language and multimodal data has surfaced.
In this study, the research team designed an innovative paradigm integrating computational modeling, behavioral experiments, and neuroscience, based on classical theories of cognitive neuroscience.
The team extracted 66 "mental dimensions" from the behavioral data of massive large models and assigned semantic labels to these dimensions. The study found that these dimensions are highly interpretable and significantly associated with neural activity patterns in the brain's category-selective regions.
The research team further compared the consistency of behavioral choice patterns across multiple models with humans. The results showed that multimodal large models perform better in terms of consistency. Additionally, the study revealed that humans tend to combine visual features and semantic information for decision-making, while large models tend to rely on semantic labels and abstract concepts.
The research indicates that large language models are not "random parrots," but rather, they have an internal understanding of real-world concepts similar to humans. The core finding is that the "mental dimensions" of artificial intelligence and humans converge on the same path.