This document describes research from the Nakayama Lab at the University of Tokyo on deep learning approaches for computer vision tasks. It discusses stacked local autocorrelation (SLAC) features, which involve iteratively computing local autocorrelations and compressing with PCA. The SLAC approach achieves better performance than standard higher-order local autocorrelation features while using lower-dimensional representations. Combining SLAC with other frameworks like Fisher vectors can further boost performance on datasets like Caltech-101. Overall, the document advocates for deep learning by stacking different single-layer modules as a simple but powerful framework.