Fig. 2: OReN architecture.
From: On the visual analytic intelligence of neural networks
a Generic architecture showing the vision model, the formation of embeddings and embedding pairs, as well the relation and decision stages. b Realization of the vision model by a CNN producing D-dimensional embeddings. c Realization of the function gθ through a four-layer MLP with N rectified linear neurons per layer. d Realization of the function fφ with a three-layer MLP with two N-sized rectified linear layers and a single linear output neuron yielding the score.
