Linear decoding performance (ridge regression, correlation of y velocities) from either latents of masked VAEs (red) or naive VAEs (blue) plotted against the mean uncertainty (i.e., returned standard deviation) of the most informative latents. Shown for different masking levels (light, all observed, to dark, 200 neurons masked), depicted in (D). Linear fits per VAE-model instantiation (seed) reveal a negative correlation between uncertainty prediction and decoding performance for the masked (slopes: -2.8 to -1.5; R-squared: 0.84 to 0.98; 10/10 seeds p<0.005), but not the naive (slopes: 0.04 to 0.22; R-squared: 0.11 to 0.87; 6/10 seeds p>0.05), approach. In gray, decoding performance directly from spikes. Note that decoding from spikes has no associated latent uncertainty.
