关键词: 非符号数量加工, 单调响应, 调谐响应, 层级编码, 除法归一化

Abstract: Numerical processing is a fundamental ability that allows organisms to extract quantity information from the environment to guide behavior, showing broad cross-species generality and emerging early in development. Recent evidence indicates that numerical processing relies not only on higher-level association areas, including the parietal and frontal cortices, but also engages early sensory regions. Here, we review the major cortical regions involved in numerosity processing and their neural response properties, with a particular focus on two characteristic response profiles, monotonic and tuned responses, and their proposed roles in numerical representation. Building on these findings, we propose a hierarchical coding framework in which divisive normalization (DN) serves as the core computational mechanism linking these two response profiles across the cortical hierarchy. DN is a canonical neural computation in which a neuron's response is scaled by the pooled activity of a surrounding normalization population, thereby regulating response gain through competitive inhibition. In early visual cortex, DN operates over spatially organized, retinotopic populations across multiple spatial scales, reducing sensitivity to non-numerical visual features such as item size while preserving population responses that vary systematically with numerosity, contributing to the monotonic numerosity signals observed in early sensory regions. As information propagates toward association cortex, the normalization pool undergoes a reorganization: rather than pooling over spatially adjacent retinotopic units, normalization increasingly operates across channels with different numerosity preferences. When a particular numerosity is presented, the channel most sensitive to that numerosity receives the strongest excitatory drive, while neighboring channels with different preferences form the normalization pool and exert competitive inhibition. This cross-channel competition sharpens the broad monotonic population response into a peaked tuned profile. Within this framework, monotonic and tuned responses are not independent coding schemes but distinct response regimes emerging from a shared normalization computation operating at different levels of the cortical hierarchy, supporting the transformation from sensory-dependent numerical signals to abstract, modality-general representations. This hierarchical normalization framework offers a unified account of numerosity-selective responses across cortical processing stages and generates testable predictions regarding the reorganization of normalization pools, changes in response selectivity, and the causal relationship between monotonic and tuned numerosity responses along the cortical hierarchy. Future research should examine whether these principles generalize across sensory modalities and symbolic numerical representations.

Key words: non-symbolic numerical processing, monotonic response, tuned response, hierarchical code, divisive normalization