Abstract
“This talk will examine the hypothesis that the role of feedback pathways in the brain is to predict, or reconstruct, patterns of feedforward activity. The matching process that this implies can be seen equivalently as readout from a content-addressable memory.
I will argue that the reconstruction of feedforward patterns by activity in feedback fibres can be done economically, provided that the receptive fields of individual cortical neurons can be described as separable products of functions along different dimensions (e.g. orientation and spatial position). This separability property seems to be true, and is central to the idea that individual cortical areas contain multiple superimposed maps for different feature dimensions. For the reconstruction to be accurate, signals from overlaid feedback axons must be combined locally in the cortex, ideally, multiplicatively, and this can be done by the apical dendrites of cortical neurons.
In addition to suggesting why apical dendrites are a ubiquitous feature of the cortex, the resulting model also explains why feedback projections are fast-conducting and have broad tangential distributions; why attention multiplicatively scales receptive field profiles; and why apical dendritic input multiplicatively scales subthreshold feedforward inputs to basal dendrites.”