This work is based on this brief outline:
- Generate synthetic data by picking a weight matrix either by hand (start with this) or randomly and running the known model forward to generate the dynamics.
- Then use this as the input to the learning system to see if it can identify the original weight matrix used to generate the observed dynamics.
- Is the learning powerful enough to be able to reverse engineer a complex dynamical system?
- How under constrained is the system (how many alternative solutions are there) that replicate the observed behavior?
- How much data does it need to constrain the problem sufficiently to identify the correct weight values?