As deep neural networks (DNNs) are increasingly deployed in critical domains, ensuring their reliability and security has become a key research focus. While mutation-based fault localization (MBFL) has shown superior effectiveness in DNN testing compared to conventional methods, its practical application faces significant computational challenges due to the enormous number of mutants it generates. To tackle this challenge, we propose DeepMR, a novel learning-based approach for mutation reduction in DNN models that maintains fault localization accuracy while dramatically decreasing computational costs. The core insight behind DeepMR is that strategically targeting key neurons—those critically influencing the decision paths of test cases—with high-impact mutation operators can yield a subset of mutants that substantially improve fault localization efficacy. DeepMR dynamically distinguishes effective mutants by jointly analyzing (1) runtime neuron behaviors during training and inference and (2) mutation operator characteristics, ensuring the selection of maximally informative mutants while discarding redundant ones. To validate the effectiveness of our approach, we trained DeepMR using data from 59 DNN programs and evaluated it on 31 DNN programs. Our experimental results show that when selecting 25% of the mutants at each layer, DeepMR can identify 90.0% of the most suspicious mutants in the neural network layers, representing a 47% improvement over random sampling. In terms of fault localization precision, selecting just 15% of mutants allows DeepMR to correctly identify all 14 bugs detected by the original MBFL without mutant reduction, while reducing the overall fault localization cost reduced by 72.3%.