SPARSE CODING OF TIME­VARYING NATURAL IMAGES

Bruno A. Olshausen
baolshausen@ucdavis.edu

We show how the principle of sparse coding may be applied to learn the forms of structure occurring in time­varying natural images. A sequence of images is described as a linear superposition of space­time func­ tions, each of which is convolved with a time­varying coefficient signal. When a sparse, independent repre­ sentation is sought over the coefficients, the basis func­ tions that emerge are space­time inseparable functions that resemble the motion­selective receptive fields of cortical simple cells. Interestingly, the coefficients form a spike­like representation of moving images, and thus suggest an interpretation of spiking activity in the brain in terms of sparse coding in time.