Figure 1
From: Tracing the Flow of Perceptual Features in an Algorithmic Brain Network
Simulation 1.
Simulation of Directed Information (DI) and Directed Feature Information (DFI). With DFI, we can establish that network node A codes and sends feature F at time t1 to B, that receives and codes F at time t2. (A) Mutual Information (MI) Curves. We simulated MEG responses from two nodes (see Methods), such that the values of stimulus feature F (e.g. the graded presence of the left eye in the face stimulus, represented with a raster of shades of blue) modulates the activity of node A at time t1, which in turn modulates the activity of node B at time t2 (node B was not directly modulated by F). The red raster plots illustrate the simulated single-trial responses of nodes A and B to values of stimulus feature F. The left and right scatters illustrate the relationship existing at time t1 (and t2) between the single-trial activity of node A (and node B, color-coded with a red raster) and the values of stimulus feature F (color-coded with a blue raster). Mutual Information time courses quantify this relationship for each time point. (B) Directed Information (DI); Directed Feature Information (DFI). For different delays, DI quantifies the total causal transfer of signal information from node A to node B-i.e. A sends MEG signal information between 100 to 140 ms to B which receives it between 210 to 230 ms–i.e. with a delay of 90 to 110 ms matching the parameters of the simulation. We schematize this relationship between sources A and B at different time points with scatters. DFI goes further by conditioning DI on the variations of feature F (represented with the conditioning performed on the different values of F, represented as different shades of blue) and so DFI quantifies how much of the total DI transfer is about feature F. Simulation 2 Simulation of Directed Information (DI) with no Directed Feature Information (DFI). We simulated MEG responses from two nodes (see Methods), such that the stimulus value independently modulated the activity of node A at time t1 and node B at time t2. The activity of B at time t2 was further modulated by the baseline activity of A at t1. MI curves and DI and DFI transfers are shown as Simulation 1. Here, there is DI corresponding the information peaks in each node, but there is no positive DFI. In this case the content of the communication is not related to the stimulus feature, it reflects instead the background activity at source A. These two functionally distinct situations cannot be dissociated based on DI and MI alone, but DFI allows us to correctly separate them.
