Vishwabrahma Shilpa, S.K. Tyagi 02 7(S)-ICCMT Based on the analysis of the error backpropagation algorithm we propose an innovative training criterion for deep neural networks to achieve maximum interval minimum classification error This criterion enhances the training process by optimizing the classification performance In addition we analyze and combine the crossentropy loss function with the Modified MultiClass CrossEntropy M3CE loss function to achieve better classification results The proposed M3CECEc model was tested on two deep learning standard databases MNIST and CIFAR10 Experimental results demonstrate that M3CE significantly improves the crossentropy loss function serving as an effective supplement to traditional crossentropy Our approach M3CECEc yielded strong performance on both datasets showing its efficacy in improving image classification tasks Copyright (c) Journal of Science Engineering Technology and Management Science. All rights reserved 2026