Fig. 3. Learning the predictive representation.

a Predictive error (L₂ norm) in blue between the network’s output and the observation as a function of the lag (Delta t). In red average L₂ norm between the observation at time 0 and at a lag Delta t. b Linear decoding of latent variables. RMS measure of the linear decoding of (x, y, θ) at time Delta t from the neural representation at time 0. The dotted line highlights the axis of symmetry of the curves. c Signal transfer analysis: Canonical Correlation Analysis between PCs of the neural representation and the latent space. The lines correspond to the average of the canonical correlations between the highlighted variables. d Same as panel c but for the observation space. e Participation ratio of the representation during learning. f Intrinsic dimensionality (ID) of the representation during learning. Five different intrinsic dimensionality estimators are used (cfr. Methods).