Diffusion MRI (dMRI) is a unique imaging technique for in vivo characterization of tissue microstructure and white matter pathways. However, its relatively long acquisition time implies greater motion artifacts when imaging, for example, infants and Parkinson's disease patients. To accelerate dMRI acquisition, we propose in this paper (i) a diffusion encoding scheme, called Slice-Interleaved Diffusion Encoding (SIDE), that interleaves each diffusion-weighted (DW) image volume with slices that are encoded with different diffusion gradients, essentially allowing the slice-undersampling of image volume associated with each diffusion gradient to significantly reduce acquisition time, and (ii) a method based on deep learning for effective reconstruction of DW images from the highly slice-undersampled data. Evaluation based on the Human Connectome Project (HCP) dataset indicates that our method can achieve a high acceleration factor of up to 6 with minimal information loss. Evaluation using dMRI data acquired with SIDE acquisition demonstrates that it is possible to accelerate the acquisition by as much as 50 folds when combined with multi-band imaging.