Inversion of V for subjects with many observations

[twoertwein]

thank you for sharing your code!

If a subject has many observations, e.g., some subjects in MPII Gaze have ~50.000 observations, you are going to invert a matrix of that size:

MeNets/scripts/em_lmecnn.m

Line 45 in 3537e99

V_hat_inv_i = V_hat_i \ eye(n_i);

Did you sub-sample your training/validation data (and were very patient at test time)? The Woodbury matrix identity can theoretically be used here but it is unfortunately numerically very unstable.