Spatial Distribution of Reconstruction Error

The position of the robot is reconstructed based on the responses from the 16 units at the highest level of the hierarchy. The resulting reconstruction error is color-coded as a function of the position within the environment. The reconstruction quality is good in large parts of the central region of the environment and becomes poorer at the borders. The latter is due to two issues: 1) the extreme border regions of the environment are not visited as often such that the estimation of the posterior probabilities becomes less accurate leading to large errors; 2) when the robot faces the wall around the border of the environment, its visual stimulus is dominated by the wall, which looks identical from different positions. This leads to perceptual singularities (i.e., same perception for different locations), yielding similar network activation patterns which can result in large reconstruction errors. (2 KB PDF)