We propose an AGREEMENT-Discrepancy-Selection (ADS) approach and aim to unify model training with sample selection by introducing antagonistic classifiers on a convolutional neural network (CNN). Minimizing the prediction gap of classifiers (by maximizing their prediction match) encourages learning CNN functions to align the distributions of tagged and unlabeled samples. Maximizing the deviation of classifiers highlights informative sampling through an entropy-based sample selection measure. The iterative prediction match gap gradually balances the distributions of the marked and unlabeled sets in a progressive distribution orientation for active learning. This work was supported in part by the National National Natural Science Foundation of China (NSFC) under the 61836012, 61771447 and 62006216 grant, the Priority Strategic Research Program of the Chinese Academy of Sciences under the XDA27010303 grant, and the China Postdoctoral Innovative Talent Support Program under the 119103S304 grant. github.com/fumengying19/AAAI21-ADS/tree/code Please consider quoting our article in your publications if the project helps your research.. .