Estimating orientation in natural scenes: A spiking neural network model of the insect central complex
Fig 3
Bump dynamics for vertical bar visual scenes.
(A) Raster plots of ring neuron and EPG cell spikes over 3 revolutions with a visual scene of a single bright vertical bar. For two revolutions both visual and angular velocity input are provided to the model. Only visual input is provided during the final revolution (green shaded region). The unwrapped ground truth heading and estimated heading (EPG cell at the center of the bump) are also shown. (B) Evolution of the synaptic weight matrix from random initial weights after 1 and 2 revolutions for the one bar visual scene. (C) Raster plots of ring neuron and EPG cell spikes over 9 revolutions. The initial 3 revolutions with a visual scene with a single bright vertical bar. For a further 3 revolutions an ambiguous visual scene with two identical bright vertical bars separated by 180 degrees (blue shaded region). For the final three revolutions the single bar visual scene was returned with the bar at one of the two offsets. A single bump of activity is maintained throughout by ring attractor dynamics. (D) Evolution of the synaptic weight matrix from random initial weights after each visual scene presentation. When two bars are presented, both positions are represented in the weight matrix. The final weight matrix can toggle between two states depending on the offset position of the final vertical bar (red box). Either maintaining the original mapping (3 revolutions vs 9 revolutions left) or remapping to the new bar position (3 revolutions vs 9 revolutions right). (E) Input current to PEN cells required to produce bump movement of different angular velocities. Below input current of 0.06 bump movement cannot be produced. (F) RMSE over the full 3 rotations when only input to PEN cells is provided for a range of angular velocities. Red star shows baseline RMSE for the angular velocity used in the 3 revolutions experiments. (G) RMSE during probe trial of the 3 revolutions experiment at a range of angular velocities. Hexagonal, Square Grid and Random symmetrical receptive field patterns all result in similar performance.