User:Birajaghoshal/sandbox

Tucker-Ghoshal Quotient (TGQ) the informativeness measurement leveraging bias-corrected estimated uncertainty selects the most informative unlabeled images as an image acquisition function in Active Learning, which significantly improves the learning process with a smaller number of required training labels. TGQ can also be used to identify mislabeled images.

This novel approach was first applied in medical image datasets for identifying both uncertain and representative unlabeled images using Bayesian Active Learning model by Biraja Ghoshal and Allan Tucker.

Instead of relying on mutual information between predictions and the model posterior, TGQ consider functional relationship and correlation among class probabilities to quantify Bias-Corrected Uncertainty (BCU) using Jackknife re-sampling method and the difference between the probability values of the highest and the second highest predictive probability value as class predictive probability distance (CPPD). This is used to calculate the Tucker-Ghoshal Quotient (TGQ) for selecting unlabeled images that are likely to be diverse and informative.

The lower the TGQ value, the higher the information content of the corresponding sample images which should represent uncertain predictions. In practice:
 * 1) TGQ -> 1 means that class predictive probability distance and uncertainty are relatively similar. This happens if a) the models have failed to reach a consensus (class membership difference is small) but model uncertainty is low, or b) the models have reached a consensus (class membership difference is large) but model uncertainty is high.
 * 2) TGQ -> 0 means that uncertainty is much larger than class membership difference. These set of images represents uncertain predictions.
 * 3) TGQ -> 1 means that uncertainty is much smaller than difference. These set represents predictions with high confidence.