These essentially embedded the MIR into tensors that we could then process on the ML side. We have not used this for any experimentation in the past year or two and the work has largely been superseded by MIR2Vec, which is what we are going to experiment with going forward.