Analyzing videos of human actions involves understanding the temporal relationships among video frames. CNNs are the current state-of-the-art methods for action recognition in videos. However, the CNN architectures currently being used have difficulty in capturing these relationships. State-of-the-art action recognition approaches rely on traditional local optical flow estimation methods to pre-compute the motion information for CNNs. Such a two-stage approach is computationally expensive, storage demanding, and not end-to-end trainable. In this paper, we present a novel CNN architecture that implicitly captures motion information. Our method is 10x faster than a two-stage approach, does not need to cache flow information, and is end-to-end trainable. Experimental results on UCF101 and HMDB51 show that it achieves competitive accuracy with the two-stage approaches.