This paper proposes an unsupervised loss function for deep semi-supervised learning that minimizes the differences in predictions of multiple passes of a training sample through a convolutional neural network (CNN). The loss function leverages the stochastic nature of techniques like dropout and randomized max-pooling to improve generalization and stability. It is combined with a mutual-exclusivity loss function to ensure predictions are mutually exclusive, enhancing model performance. The method is evaluated on several benchmark datasets including MNIST, SVHN, NORB, CIFAR10, CIFAR100, and ImageNet. Results show that the proposed method significantly improves classification accuracy, especially when only a small number of labeled samples are available. The approach is implemented using two CNN frameworks, cuda-convnet and sparse convolutional networks, and demonstrates effectiveness across various tasks and data sizes. The combination of the transformation/stability loss and mutual-exclusivity loss leads to further improvements in model accuracy. The method is shown to be effective in both supervised and semi-supervised learning scenarios, with the proposed loss function being fully unsupervised and compatible with any supervised loss function. The results indicate that the proposed method achieves state-of-the-art performance on multiple benchmark datasets.This paper proposes an unsupervised loss function for deep semi-supervised learning that minimizes the differences in predictions of multiple passes of a training sample through a convolutional neural network (CNN). The loss function leverages the stochastic nature of techniques like dropout and randomized max-pooling to improve generalization and stability. It is combined with a mutual-exclusivity loss function to ensure predictions are mutually exclusive, enhancing model performance. The method is evaluated on several benchmark datasets including MNIST, SVHN, NORB, CIFAR10, CIFAR100, and ImageNet. Results show that the proposed method significantly improves classification accuracy, especially when only a small number of labeled samples are available. The approach is implemented using two CNN frameworks, cuda-convnet and sparse convolutional networks, and demonstrates effectiveness across various tasks and data sizes. The combination of the transformation/stability loss and mutual-exclusivity loss leads to further improvements in model accuracy. The method is shown to be effective in both supervised and semi-supervised learning scenarios, with the proposed loss function being fully unsupervised and compatible with any supervised loss function. The results indicate that the proposed method achieves state-of-the-art performance on multiple benchmark datasets.