Figure 4: Robustness to naturally occurring recording condition changes.

We created decoder evaluation conditions in which the neural inputs were likely to be different from much of the training data by withholding access to the most recent several months of data. Each circle corresponds to the mean closed-loop BMI performance using these ‘stale’ MRNN (red), FIT Long (dark blue) and FIT Old (teal) decoders when evaluated on six different experiment days spanning 7 (13) days in monkey R (L). Each test day, these three decoders, as well as a FIT Sameday decoder trained from that day’s arm reaches, were evaluated in an interleaved block design. The legend bars also denote the time periods from which training data for each stale decoder came from. We repeated the experiments for a second set of decoders to reduce the chance that the results were particular to the specific training data gap chosen. The training data periods contained 82 and 92 data sets (monkey R) and 189 and 200 training data sets (monkey L). The only decoder that was consistently usable, that is, did not fail on any test days, was the MRNN. To aid the interpretation of these stale decoder performances, we show the average performance across the six experiment days using arm control (grey dashed line) or a FIT Sameday decoder (blue dashed line).