170f6aa36530c364b77ddf83a84e7351-AuthorFeedback.pdf

Note that the assumption on encoder gap is very mild. Intuitively, if a dictionaryD provides quickly decaying9 approximation error as a function of the cardinality, then a positive encoder gap exists for somes. Lastly,while our contribution is mostly the-18 oretical, we also provide further numerical19 evidence for the encoder gap in real scenar-20 ios. Wetrainedtwomodels,onSVHN(with21 256 atoms) and CIFAR (with 1024 atoms),22 with thetraining procedure described inour23 paper, and depict the value ofτs (obtained24 over a collection of samples), for different25 values ofλ. Deploying sparse priors in28 the learned representations is a first-take at the analysis of non-linear and data-dependent mappings for supervised29 models.