SINGLE-CELL PROPERTIES DETERMINE REPRESENTATIONAL GEOMETRY OF CORTICAL LAYERS

Recent studies have moved us away from thinking about single cell sensory/motor representations and towards the activity of whole networks. What defines the geometry of network representations? While network connectivity can impact representational geometry, it is unclear to what degree geometry is also impacted by single cell properties. For instance, the neocortex is organized into distinct layers, where the cells in each layer have unique physiological properties. We investigated whether these single-cell properties may change the dimensionality and geometry of representations at the population level. Firstly, we measured how cells transform their current input into action potentials – the frequency-current (fI) curve. To estimate neural activity under physiological conditions, we recorded in awake head-fixed mice via whole-cell patch clamp (total N=80 neurons). We found differences among the cortical layers. In particular, we found that pyramidal neurons in Layer 2/3 have a higher threshold current and lower firing rates than those in Layer 5 (Panel A). Using abstractions of the recorded fI curves as activation functions in artificial neural networks, we tested the resulting representational geometry (Panel B). We trained the networks on classification tasks, modelling discrimination behaviours that are thought to require early sensory areas. Preliminary results suggest that networks with L5-like functions better represent the precise patterns of inputs, whereas networks with L2/3-like functions better encode global stimulus class. Overall, we conclude that the fI curve of single-cell types can impact network representations within cortical layers, which may allow the layers to specialise during sensory discrimination in vivo.