<p>Deep learning is a highly popular and effective classification technique capable of handling complex patterns. In recent years, many researchers have focused on enhancing the performance of both shallow and deep intelligent classifiers. Among the various methodologies developed for this purpose, the reliable and jumping modeling techniques have shown great promise in improving the accuracy of diverse classifiers with different characteristics. In this paper, a reliable jumping-based deep learning (RJDL) approach is proposed that simultaneously leverages the advantages of these methodologies to enhance the classification performance of deep learning classifiers. In the first stage of the RJDL classifier, the jumping-based methodology is first applied to the cost/loss function of the conventional deep learning classifiers. This transformation converts the continuous feasible set into a discrete one, enabling the deep learning model to jump between different possible points. In the second stage, the reliable-based methodology is then employed on the jumping-based cost/loss function obtained from the previous stage. This approach helps to estimate the discrete connection weights in a manner that the frequency of jumping is minimized. To evaluate the performance of the proposed RJDL methodology, seven benchmark data sets related to the transportation is considered. Additionally, for the implementation of the proposed methodology, the deep feed-forward neural networks (DNNs) is selected. Empirical results of the reliable jumping-based deep feed-forward neural network (RJDFNN) demonstrate that the proposed classifier consistently yields more accurate outcomes compared to the conventional deep feed-forward neural network. On average, the RJDFNN classifier achieves a classification rate of 89.87%, which is 6.65% higher than the classic deep feed-forward neural network.</p>