Variance is also another heuristic that you can use to rank the images to be labeled.
To put it simply, if the variance of the score of an image is low among the different classes, then it might suggest that the image is more "informative". So, it makes sense to label this image for the model to progress. It works because a low variance in the classification score implies that the model is not really certain about the prediction. To help the model fix such an issue, we label this data asset and let the algorithm learn from this example.
Learn more about the other heuristics:
Only 13% of vision AI projects make it to production, with Hasty we boost that number to 100%.