In a recent paper about progress in computer animation a so called motion graph is used to describe the transition between keyframes of facial animation. Easy Generation of Facial Animation Using Motion Graphs, 2018 As far as i understand from the paper, they used a motion capture device to record faces of real people and extract keyframes. Then a transition matrix was created to ensure that a walk from keyframe #10 to #24 is possible but a transition from keyframe #22 to #99 is forbidden.
The idea itself sounds reasonable good, because now a solver can search in the motion graph to bring the system from a laughing face to a bored face without interruption or unnatural in-between-keyframes. But wouldn't it be great if the transition matrix can be stored inside a neural network? As far as i understand the backpropagation algorithm , the neural network can learn input-output relations. So the neural network has to learn the transition probability between two keyframes. And a second neural network can then produce the motion plan which is also be trained by a large corpus. Is that idea possible or is it the wrong direction?