Fig. 7: Perception and memory responses can be simulated with a bidirectional hierarchical model.
From: Perception and memory have distinct spatial tuning properties in human visual cortex
a Illustration of stimulus-driven activity propagating through a linear hierarchical network model in the feedforward direction (left) and mnemonic activity propagating through the model in the feedback direction (right). In both cases, a given layer's activity is generated by convolving the previously active layer's activity with a fixed Gaussian kernel. The feedforward simulation began with a boxcar stimulus. The feedback simulation began with duplication of the feedforward activity from the final layer. b Example of feedforward and feedback simulations for one set of parameters (stimulus = 15∘; kernel σ = 15∘, number of layers = 8), plotted in the conventions of Fig. 4b. The feedforward simulation parallels our observations during perception, and the feedback simulation parallels our observations during memory. c Location, amplitude, and FWHM parameters for each layer, plotted separately for feedforward and feedback simulations in b. Location is preserved across layers in the feedforward and feedback direction. Note that FWHM becomes progressively wider in later layers in the feedforward direction and in earlier layers in the feedback direction. This results in large differences in FWHM between feedforward and feedback activity in early layers. These trends closely follow our observations in Fig. 4c. See Supplementary Fig. 7 for other simulations with different layer numbers, stimulus sizes, and kernel sizes.
