cmlnn

Learning to use spatiotemporal chaos

Basic idea

Coupled map lattices (CMLs) exhibit robust nonlinear functional capacity with few parameters. As such they might prove a useful computational tool for deep neural networks. Currently this is just a first-pass notebook.

Errata

• This is in Keras but must move to PyTorch:
  • I need layerwise learning rates so that dynamics can be learned slower than initialization, which Keras doesn't yet support
• Learning requires a slight modification of CML update:
  • the lattice boundaries must update independently instead of referencing neighbors.
    • visually, dynamics seem unaffected by this, but should validate further