Modify de_StatsFreqPreferences and de_PlotFreqPreferences to examine frequency preferences on units trained on natural images.
Current results:
Plot focused on below selects the max (mean = mean response, std = variability in response) over secondary dimensions on the primary dimension. So, for freq, large bars mean higher responsiveness to variation in orientation and phase.
I expected that, to be biased for a specific frequency, you should have higher variability for that frequency - to better distinguish variations (in orientation & phase) at that frequency.
Preferences in all sigmas for vertical / horizontal (not surprising)
Interaction between sigma and frequencies (in both mean, std).. but in the opposite of the expected direction - lower sigmas show higher responsiveness to higher, not lower, frequencies.
vishaalprasad
commented
8 years ago
Modify
de_StatsFreqPreferencesandde_PlotFreqPreferencesto examine frequency preferences on units trained on natural images.Current results:
Plot focused on below selects the max (mean = mean response, std = variability in response) over secondary dimensions on the primary dimension. So, for freq, large bars mean higher responsiveness to variation in orientation and phase.
I expected that, to be biased for a specific frequency, you should have higher variability for that frequency - to better distinguish variations (in orientation & phase) at that frequency.