Why the cycle loss work

[henrylovedesign]

Hi, the autoencoder loss and cycle loss described in the paper should has a global optimal solution that is just copying quaternions and velocity, but why the GRU model learn something different?

[rubenvillegas]

1. The method is fully unsupervised in the rotations output. In addition, there is only an objective encouraging joint positions. Since the rotation are unsupervised, you can reach those positions by rotating bones in any way you want as long as they are reached. Therefore, it not guaranteed that it will learn to copy the rotations.

2. The velocities are learned by the neural network in a representation where the rotations are also entangled. We don't encourage the neural network to disentangle some neurons to learn to only the velocity factors. Therefore, it is not guaranteed that copy will be learned. It will learn estimates from which tiny errors can accumulate over time and cause large deviations from the true trajectory.

I hope this helps.