With regards to this plot from 2D classification:
I assumed that “current sigma” meant the sigma at the current iteration, while “initial sigma” would correspond to the sigma at the start of classification.
However, the initial and current sigma are not the same at iteration 0, suggesting that this is not the case. How is the initial sigma defined?
Here’s a bit of clarification on what these values are:
- Initial sigma is, as you say, the sigma (standard deviation of all Fourier components, rotationally averaged) at the start of classification. This value does not take into account the projections and uses 1000 random images.
- The current noise is the standard deviation of all Fourier components of the data minus the projection. Put another way, current noise can be thought of as the noise “left over” after we subtract our 2D classes. This parameter is calculated at every iteration using the updated class averages, and using all particles in the minibatch.
- Finally, the current sigma is the noise model that was actually used during the classification. So this includes things like the uncertainty factor and the annealing between initial sigma (or white noise) and the current noise.
Here’s a comparison of current sigma across iterations with the default settings (start annealing after iteration 2, anneal over 15 iterations (so, annealed at 18):
I hope that’s helpful!