Numerical cognition is a fundamental skill for human being, underlying everyday activities such as paying bills as well as complex achievements like engineering and scientific discovery. Numerical information is inherently multi-sensory: whether we see five pens on a table, hear five knocks on a door, or feel five taps of a hammer, we are able to estimate their numerosity. This thesis begins from this multi-sensory nature of numerosity to explore how the brain represents and develops numerical information. In the first chapter, we investigated how cross-modal (i.e., vision vs. audition) and cross-format (i.e., sequential vs. simultaneous) numerosity are represented and converge in the brain. Using functional magnetic resonance imaging (fMRI) combined with multivariate pattern analysis (MVPA), we examine the fine-grained organization of neural representations of numerical information. These analyses reveal how distributed activity patterns across sensory and parietal regions encode numerosity and how such codes may generalize, or remain distinct, across sensory and temporal dimensions. In the second chapter, we challenge a long- standing assumption in the literature: that numerosity is inherently tied to vision. Drawing on behavioral, psychophysical, and neuroimaging evidence, we argue that visual experience is not a prerequisite for developing numerical cognition. Together, this thesis advances our understanding of how the human brain encodes numerical information across sensory modalities and formats.