Fig. 4. Comparison between predictive and non-predictive learning.

We train 50 networks of 100 neurons in each of the predictive and non-predictive conditions and equalize the learning axis between the two to highlight the trends of the different measures. a Predictive error. The position of the predictive error symmetry axis plotted throughout learning for the predictive and non-predictive network ensembles. The symmetry axis position is the one that minimizes a L2 norm between the predictive error curve (cf. Fig. 3a) and its reflection through the symmetry axis. b Latent signal transfer analysis. A canonical correlation analysis is performed between the latent space and the top PCs of the neural representation at every epoch, and the average of the two canonical correlations (for coordinates x and y) is shown. c Observation signal transfer analysis. The canonical correlation analysis, same as panel b, is performed between the top PCs of the observations and the top PCs of the network’s representation. d Linear dimensionality (PR) throughout learning. e Non-linear dimensionality (ID) throughout learning. f Dimensionality gain (DG) throughout learning.