Por Christian Machens (Fundação Champalimaud, Lisbon, Portugal).

Models of neural networks can be largely divided into two camps. On one end are functional models, such as rate networks, that can perform a multitude of functions and have led to many recent breakthroughs in ML/AI, but ignore well-established biological facts. On the other end are mechanistic models such as balanced spiking networks that resemble neural activity, but are limited to simple computations. Here, I will introduce a new framework for spiking networks which retains key properties of both mechanistic and functional models. The key insight is to recast the problem of spiking dynamics in a lower-dimensional space of network activity modes rather than in the original neural space. I will illustrate these insights with simple, geometric toy models, and show how they allow us to construct networks that are both computationally powerful, while reproducing key biological facts. I will argue that these results force us to reconsider the very basics of how we think about neural networks.