Researchers at the Institute of Automation, Chinese Academy of Sciences, have introduced a groundbreaking neural network framework known as CATS Net. This development addresses a fundamental limitation in current artificial intelligence: the inability to replicate the human capacity to abstract concepts from sensory experiences and think within a conceptual space. While human cognition relies on compressing high-dimensional sensory input into low-dimensional concepts and then reconstructing perception from these concepts, AI systems have historically struggled to perform this bidirectional process. CATS Net successfully bridges this gap by simulating human-like concept formation, understanding, and communication. The framework consists of two primary components: a Concept Abstraction (CA) module and a Task Solving (TS) module. When processing visual tasks, the CA module automatically compresses high-dimensional visual inputs into compact, low-dimensional "concept vectors." These vectors function analogously to keys in a lock. Through a hierarchical gating mechanism, they generate a series of switching signals that dynamically regulate the neural activity within the TS module. This process efficiently and flexibly guides the system to complete specific visual perception tasks. A unique feature of CATS Net is its ability to autonomously generate a vast array of new concepts through interaction with its environment, thereby forming a personal concept space. When different neural networks achieve alignment in their concept spaces, knowledge can be transferred directly via concept vectors, mirroring how humans share ideas. To validate the biological plausibility of their model, the research team compared the spontaneously formed concept representations in CATS Net with human conceptual spaces and neural activity data. Representational similarity analysis using functional magnetic resonance imaging (fMRI) revealed that the concept space generated by CATS Net aligns with established psychological models of human semantic cognition. Furthermore, its representation patterns corresponded with activity in the ventral occipitotemporal cortex of the human brain, the region responsible for visual semantic representation. Additionally, the dynamic gating mechanism of the CA module matched the activity patterns of the semantic control network in the brain, which is involved in concept extraction and manipulation. These findings indicate that CATS Net not only functionally simulates human concept cognition but also elucidates the computational principles behind the human brain's concept formation and understanding at a mechanistic level. This work provides a significant foundation for developing next-generation intelligent systems capable of human-like concept application and formation. The research results were published online in the journal Nature Computational Science. The project was supported by the National Natural Science Foundation of China and various relevant projects from the Chinese Academy of Sciences. This achievement marks a pivotal step toward building artificial intelligence systems that truly understand and communicate through the lens of human-like conceptual thought.