Large-scale image recognition via unsupervised feature learning

In this talk I will share our experience about large-scale image recognition by using feature learning. We worked on extending sparse coding to a broader family of nonlinear coding methods that explore the geometrical structure of sensory image data. The coding of image local features gives rise to significantly better features, which enables simple linear classifiers to produce stronger results, and also scale much better, than nonlinear SVMs using Chi-square or intersection kernels. The methods achieved state-of-the-art results on a range of challenging image classification tasks, including Caltech 101, Caltech 256, PASCAL VOC, and ImageNet.