Supplementary Figure 6: Dynamically stitched multisession LFADS model outperforms single-session LFADS models in predicting reaction times.

As defined in (Kaufman et al., eNeuro 2016), the condition-independent signal (CIS) is a high variance component of motor cortical population activity obtained via demixed principal components analysis (dPCA). Kaufman et al. 2016 also demonstrated that threshold crossing time of the CIS on single trials is an effective predictor of reach reaction time (RT). Here we identify the CIS as a linear projection of LFADS factor trajectories. We apply dPCA to the factor outputs of each single-session and the multi-session LFADS models to identify the largest condition-independent component, and then threshold the CIS to predict RT on single trials. a. Plot of condition-independent signals (CIS) for an example dataset. Each trace represents the CIS timecourse on a single trial, and is colored by that trial’s actual RT. b. Plot of correlations between CIS-predicted RT and actual RT on trials from each dataset for stitched multi-session LFADS vs single-session LFADS. Each point represents an individual recording session. For the stitched model, a single CIS projection was computed and applied for all sessions, whereas individual CIS projections were obtained for each single-session model.