Opening the black box of neural variability: From noise to mechanisms

Neural models typically emphasize dominant trends in data while treating the remaining variability as random noise. In biological neural systems, however, variability plays more nuanced and multifaceted roles. Although variability arising from noise sources can hinder functionality, biological systems have evolved to incorporate it in ways that enhance robustness, while flexibly enabling diversity in structure and function, both within individuals and across populations. This reduces the genetic instructions needed to construct and operate neural systems, compared to a uniquely prescribed building plan. These benefits are often overlooked in computational modeling. Here, we advocate for modeling approaches that replace unmodeled variability with mechanisms supporting efficiency, robustness, and flexibility. Such strategies have successfully replaced the notion of noisy and unreliable sensory neurons with a principled account of their activity. Similar advances are ongoing in the study of brain development. We discuss studies on neural growth, plasticity, and self-organization indicating that variability should not be treated as a nuisance, but as a means of discovering key operating principles of neural systems. Generative models, which are focused on the principles underlying variability, could play a key role in advancing theoretical neuroscience.