journal menu
![[Figure 3]](sor5003fig3mag.jpg)
|
|
Figure 3
Jittering regularizes InstaMap. Volume resolution at FSC0.5 as a function of the number of particles used for training for different jittering scaling values J. The loss is shown as insets for the training and validation sets. With few (<1000) particles and many (>20) training epochs, InstaMap fits high-resolution noise. This can be ameliorated by jittering the grid and increasing its scaling J. This acts in a similar way as a low-pass filter, preventing both high-resolution signal and noise overfitting, as can be seen from the light-blue (J = 0.05) and dark-blue (J = 10) panels on the right for two poses. For each example pose, the four subpanels from top to bottom show the top view of the TRPV1 ion channel: projected reference volume, synthetic data point, projected InstaMap with CTF and projected InstaMap. These are shown for increasing InstaMap training time points at gradient steps 47, 94, 522, 4985 and 9495. |
![[Figure 3]](sor5003fig3.jpg)
 | STRUCTURAL BIOLOGY |
ISSN: 2059-7983
access
