Brain image registration transforms a pair of images into one system with the matched imaging contents, which is of essential importance for brain image analysis. This paper presents a novel framework for unsupervised 3D brain image registration by capturing the feature-level transformation relationships between the unaligned image and reference image. To achieve this, we develop a feature-level probabilistic model to provide the direct regularization to the hidden layers of two deep convolutional neural networks, which are constructed from two input images. This model design is developed into multiple layers of these two networks to capture the transformation relationships at different levels. We employ two common benchmark datasets for 3D brain image registration and perform various experiments to evaluate our method. Experimental results show that our method clearly outperforms state-of-the-art methods on both benchmark datasets by a large margin.