Figure 9
From: Neural network based successor representations to form cognitive maps of space and language
MDS of the word successor representation vectors: a two-dimensional projection of the 40-dimensional word SR vectors (rows in SR matrix) for the calculated ground truth (A) and the learned SR matrix (B). In both cases, the words cluster according to the respective word class. Different colors correspond to word classes again. However, the resulting clusters are less dense and located closer to each other than for the TP vectors. Since SR vectors cover several time steps, whereas TP vectors only cover a single time step in the future, this result is intuitive. Note that, scaling of the axes is in arbitrary units since coordinates have no particular meaning other than indicating the relative positions of the projected vectors.
