Figure 3.

Extracting latent factors from the V4 and PFC population activity. A, Schematic of dimensionality reduction with GPFA. A low-dimensional subspace of the neural population activity space was identified, which captured shared variability of activity among neurons within each brain area (left), resulting in a lower dimensional set of latent factor scores (right). B, Average factor scores for one factor for each brain area from a representative session. At each time point (e.g., the dashed line at −10 ms relative to stimulus onset; insets show distributions across trials of factor scores at the indicated time), we trained a logistic regression decoder to classify trials as cue-in-RF or cue-away using the factor scores. One factor is shown for illustrative purposes; in actuality 10 factors were used for decoding. The gray rectangle indicates the time when the stimulus was present (400 ms duration). Shading represents ±1 SEM.