Across brain regions and species, one key feature of neural activity is that responses are highly variable. Hence, (one of) the biggest computation problems of the brain is to compensate for its own internal noise. This interpretation is challenged by experimental data: in many contexts the brain seems to actively put itself in a dynamic regime where responses are highly variable, which suggests that there may be computational advantages to having a seemingly ‘noisy’ brain. In this talk I will discuss a new theoretical framework for how low-dimensional structured noise can be used to dynamically route task-specific information between neural populations. I will show how appropriate noise structure can be learned in artificial neural networks from limited data and find signatures of such coding in population recordings from macaque V1 and MT during a discrimination task (Ruff & Cohen, 2016). <a href="https://as.nyu.edu/content/nyu-as/as/faculty/cristina-savin.html">Learn more</a>|full_html