FIGURE 7 Filter learned in the first layer on the CIFAR-10 dataset by traditional CAE (no pooling and 0% noise) and denoising CAE (no pooling and 50% Gaussian noise)
4.1.3 | Convolutional Sparse coding
Sparse coding is a group of biologically plausible algorithms that leans the basis function to capture the high-level features from the unlabelled dataset. In Sparse coding, a sparse vector is computed through the linear operation with a learned dictionary matrix for the optimum input reconstruction. Sparse coding was compared with batch alternatives and claimed to be an efficient feature learner over a large dataset. However, the traditional sparse coding is computationally expensive as it is performed over the whole image, and the representations are often redundant as inferences are performed at the patch level.