Towards a more general understanding of the algorithmic utility of recurrent connections
Fig 5
(A) Randomly generated examples of the competitive foraging task. (B) The animal’s and competitors’ propagation networks. Each one implements the tag-propagation algorithm with open space pixel corresponding to on pixels and barriers corresponding to off pixels. Unlike the edge-connected pixel task, the source pixels change in every sample to correspond to the location of the animal and its competitors respectively. See Methods for further implementation details. (C) Architecture for the trained decision network. The time series of animal and competitor ranges are concatenated with the food pixel and then are used as the input to a recurrent network. The decision traces shown in panel E are the projection of the trained network onto the two-dimensional readout. (D) Sample outputs for the trained propagation networks. These are the best networks trained with 7, 10, and 12 layers respectively. Errors in the trained network are marked in red. We show results only for the competitors’ network as the propagation task is the same for both the animal and its competitors; only the input which corresponds to the initial location changes. (E) The generalized tag propagation implements a correct version of the decision trace and is described in the Methods. For comparison, we show the decision trace outputted by the trained decision network. Propagation is shown after ten time steps, and each example is labelled with the correct decision after ten steps: “stay,” “run,” or “out of range.” Note that if the food location is in the range of both groups of animals, the decision is based on which animal can reach the food first.