June 11, 2011 - 1AI has learned from theChinese Academy of SciencesThe WeChat public number of the Institute of Automation was informed that recently a joint team of the Neurocomputing and Brain-Computer Interaction (NeuBCI) group of the Institute and the Center of Excellence in Brain Science and Intelligent Technology Innovation of the Chinese Academy of Sciences (CAS) combined behavioral experiments and neuroimaging analyses toFirst demonstration of multimodalLarge Language Model（MLLMs) The ability to spontaneously develop a conceptual representation system for objects that is highly similar to that of humansThis study not only opens a new path for the cognitive science of artificial intelligence, but also provides a theoretical framework for constructing artificial intelligence systems with human-like cognitive structures. This research not only opens up a new path for the cognitive science of artificial intelligence, but also provides a theoretical framework for the construction of artificial intelligence systems with human-like cognitive structures. The research results are published in Nature Machine Intelligence under the title of Human-like object concept representations emerge naturally in multimodal large language models. Intelligence.)
The cognitive ability of humans to conceptualize objects in nature has long been regarded as the core of human intelligence. When we see a "dog," "car," or "apple," we not only recognize their physical characteristics (size, color, shape, etc.), but also understand their function, emotional value, and cultural significance -- this multidimensional conceptual representation forms the cornerstone of human cognition. This multidimensional conceptual representation forms the cornerstone of human cognition.
Traditional AI research has focused on the accuracy of object recognition, but little has been done to explore whether the model actually "understands" the meaning of the object. Researcher He Huiguang, corresponding author of the paper, pointed out, "Current AI can distinguish between pictures of cats and dogs, but the essential difference between this 'recognition' and human 'understanding' of cats and dogs remains to be revealed." Starting from the classical theories of cognitive neuroscience, the team designed a set of innovative paradigms integrating computational modeling, behavioral experiments and brain science. The study used the classic cognitive psychology "triplet odd-one-out" task, in which a large model and humans are asked to select the least similar option from a triad of object concepts (any combination of 1,854 everyday concepts). By analyzing 4.7 million behavioral judgments, the team constructed the first-ever "concept map" of an AI macromodel.
The researchers extracted 66 "dimensions of mind" from a large amount of large model behavioral data and assigned semantic labels to these dimensions. The study found that these dimensions are highly interpretable.and was significantly correlated with patterns of neural activity in category-selective areas of the brain (e.g., FFA for processing faces, PPA for processing scenes, and EBA for processing torsos).
The study also compared the consistency of multiple models in behavioral choice patterns with humans (Human consistency). The results showed that multimodal large models (e.g., Gemini_Pro_Vision, Qwen2_VL) performed better in terms of consistency. In addition, the study reveals that humans are more inclined to combine visual features and semantic information to make judgments when making decisions, whereas grand models tend to rely on semantic labels and abstract concepts.This study shows that large language models are not "random parrots" and that there is a human-like understanding of real-world concepts within them.