Hi,
In some of our initial tests with 3D-Flex, we are seeing linear streaks in the latent distribution (see example below).
Is this normal? It smells like overfitting of some kind. If so, what should we tweak to fix?
Cheers
Oli
Same here
Hey @ehanssen @olibclarke – this ‘Moiré’-esque pattern is a side effect of the coordinate descent on the latent space within the auto-decoder stage (cf. the ‘Auto-decoder’ section of the preprint).
The coordinate descent is implemented via an independent search over a grid in each latent dimension, so the particles can only fall into discrete locations. This discretization is probably more/less visible depending on matplotlib settings and particle density.
Finally, as described in the paper, please note that these discrete latent coordinates are also perturbed by some noise during training to encourage smooth motion when moving through the latent space. We do not visualize this noise injection in this plot.
2 Likes
Thanks @vperetroukhin - what about the radial streaks in the distribution? Are these expected, or a sign of overfitting?
Just had a quick discussion about this – we’re not sure what those are! You could try to increase Noise injection stdev in the training job to see if that dithers those streaks.
1 Like