Retrieval-centric deep learning explores replacing weight matrices with growing databases

A researcher describes retrieval-centric deep learning as a framework for replacing neural-network weight matrices with growing databases of training inputs and backpropagated gradients. The team reports that its layers matched but did not outperform strong image-classification baselines, at substantially higher cost. In unpublished synthetic-language experiments, the researcher says growing RBF layers and approximations sped up fact acquisition, but the large widths needed for significant gains remain too expensive for use at scale.