Motion-compensated reconstruction has shown to be effective for suppressing imaging noise and reducing motion blur in cardiac-gated SPECT imaging. In this work we investigate the feasibility of using a cascaded learning network for motion correction in a sequence of gated images acquired at different phases of the cardiac cycle, which are known to suffer from both imaging degradation and large-extent motion deformation. To be realistic, we make use of a set of consecutive clinical acquisitions (from 130 subjects) which includes both normal and abnormal subjects. The quantitative results in the experiments indicate that the learning network can yield consistently more accurate compensation results than the classical optical flow equation (OFE) method both on standard dose and half dose studies. In particular, the learning network achieved a relative MSE of 0.021 (full dose) and 0.037 (half dose), compared to 0.035 (full dose) and 0.053 (half dose) for OFE.