Fig. 3: Visual analytic reasoning approach with a saccadic network.
From: On the visual analytic intelligence of neural networks
a The recurrent units receive D-dimensional image embeddings and eye-position inputs from a series of simulated saccadic eye movements and perform reasoning using the internal temporal dynamics. After SI initial saccades, the final decision is determined from the outputs of the final sigmoidal (σ) neuron over SE evaluation saccades. b The neural dynamics of biological neurons are incorporated into a deep network through SNUs, where each SNU represents an abstraction of the LIF model that integrates inputs from synapses into its membrane potential Vm, and generates an output spike each time Vm crosses the spiking threshold Vth.
