Animations of figures, labelled according to the corresponding figure
and panel in the paper (animated GIFs, view with internet browser or
with xanim/gifview in Linux):

- fig2GG1: animation of f(x) and X for RPM.
  Parameters: lf = 0.1; lF = 0; sx = 0.08.
  Note how the lack of a function F in RPM (unlike DRUR) causes X to
  split into several chunks (panel G1, right) and to blow up folds
  present in the initial X from Laplacian eigenmaps (panel G, left).

- fig2GG1-other: animation of f(x) and X for RPM, with different
  parameters from those in the corresponding panels in the paper:
  lf = 1; lF = 0; sx = 0.08.
  Note how the lack of a function F in RPM (unlike DRUR) causes X to
  split into 2 chunks (panel G1, right) and to blow up a fold present in
  the initial X from Laplacian eigenmaps (panel G, left).

- fig3B: animation of the latent space (X and F(Y)) for DRUR with
  the mocap data over training iterations.
  Parameters: lf = 0.8; lF = 0.5; sx = 0.8; sy = 0.8.

- fig3E-blue.gif, fig3E-green.gif, fig3E-red.gif: animation of how DRUR
  smoothly and realistically reconstructs a running human pose (by
  following a continuous path in latent space). Each animation
  contains 100 frames, none of which are in the training set.

- fig4AB: animation of the latent space (X and F(Y)) for DRUR with
  the faces data over training iterations.
  Parameters: lf = 0.2; lF = 0.2; sx = 0.3; sy = 0.3.

- digit2.gif: like fig4AB but for the digit-2 data.
  Parameters: lf = 0.2; lF = 0.5; sx = 0.3; sy = 0.8.
  In all these figures (particularly for the faces & digits) note how
  DRUR drastically improves the initial LE embedding.