Here is a new video demonstrating the generation of sequences of motor patterns learnt from human motion capture data, using a single neural network under the control of conceptors.
Generating a sequence of motor patterns learnt from human motion capture data, using a single recurrent "reservoir" neural network under the control of conceptors (mocap data retrieved from the CMU motion capture repositoryhttp://mocap.cs.cmu.edu/; visualization generated with the MoCap toolbox from the Department of Music at the University of Jvaskylan
https://www.jyu.fi/hum/laitokset/musiikki/en/research/coe/materials/mocaptoolbox). The conceptors soft-constrain the network states to certain subspaces (notice the color bars in the left bottom of the video: they show the activation of conceptors; this activation sequence is the only control input needed to produce the behavior sequence). Soft transitions between behaviors are obtained automatically from the network dynamics.