Fig. 6: Predictive representations of arm-reaching movements. | Nature Communications

Fig. 6: Predictive representations of arm-reaching movements.

From: Predictive learning as a network mechanism for extracting low-dimensional latent space representations

Fig. 6: Predictive representations of arm-reaching movements.

a Plane transverse to the dynamic of arm-reaching movements. The muscle model is shown and the two latent angular variables α and β. b Recurrent network model. c Predictive error upon training. The symmetry axis is around lag +1 indicating that the network is carrying out the prediction correctly. d Latent signal transfer and observation signal transfer. e Dimensionality trends across learning for both linear (PR) and nonlinear (ID) dimensionality measures. f Top: principal components space (PCs 1–2) colored by the average angles α, β for each location. Bottom: average activity of neurons in the space spanned by the top 2 PCs. Each subplot represents the average activity of a single neurons. Neurons are ranked according to their average firing rates. The most active neuron is in the top left corner, the second in the first column second row and so on for all the neurons. g Average activity of neurons in latent space α, β. Each subplot corresponds to the neuron in panel f.

Back to article page