Figure 1.

Overview of the response models. (A) Response models in the RNN family. All RNN models process feature sequences via two (recurrent) nonlinear layers and one (nonrecurrent) linear layer but differ in the type and number of artificial neurons. L-10/50/10 models have 10, 50, or 100 long short-term memory units in both of their hidden layers, respectively. Similarly, G-10/50/10 models have 10, 50, or 100 gated recurrent units in both of their hidden layers, respectively. (B) First-layer long short-term memory and gated recurrent units. Squares indicate linear combination and nonlinearity. Circles indicate elementwise operations. Gates in the units control the information flow between the time points. (C) Response models in the ridge regression family. All ridge regression models process feature sequences via one (nonrecurrent) linear layer but differ in how they account for the hemodynamic delay. R-C(TD) models convolve the feature sequence with the canonical hemodynamic response function (and its time and dispersion derivatives). R-F model lags the feature sequence for 3, 4, 5, and 6 s and concatenates the lagged sequences.