Figure 2.

Stable maintenance occurs in a “mnemonic” subspace despite heterogenous dynamics of individual units. A, Projection of input stimulus patterns in the mnemonic subspace. “Stimulus PC1” and “Stimulus PC2” correspond to the two PC dimensions with the largest variance across time-averaged representations of the mnemonic coding st-RNN hidden layer units. Points with lighter shades of gray represent input patterns of lower sparsity (proportion of active pixels). Insets, Specific 8 × 8 input patterns. B, Top row, Activity dynamics of the top 16 hidden layer units with the most “stable” dynamics (see text for details) in the mnemonic coding st-RNN during the blank epoch (t = 4–20); each color denotes dynamics for a different unit. Dots: data for every third time bin; lines: data for each time bin. Bottom row, Same as in top row but for the top 16 hidden layer units with the most “unstable” dynamics. Left and right columns, Results for two exemplar input patterns. Several units show overlapping activity profiles. C, Left, Low-dimensional trajectories obtained by projecting the hidden layer unit activity onto the mnemonic subspace during the maintenance epoch. Each cluster of points corresponds to a different input 8 × 8 pattern (insets). Blue to yellow: time points ranging from early (t = 3) to late (t = 17) in the blank epoch. Right, Same as in the left panel, but trajectories plotted with an additional dimension along the z-axis indicating a “Time PC” corresponding to a PC dimension with maximal variance in input-averaged activity across time that is also orthogonal to the mnemonic subspace. Dark-gray trajectories on the xy-plane indicate the projection of the trajectories in the mnemonic subspace.