Credit assignment

Backpropagation remains the gold standard, but alternative approaches that don't approximate it have been recently gaining some steam.

1. Autodiff implementations in deep learning libraries - basically the chain rule
2. Biologically plausible implementations
   1. Multiplexing
   1. Burst-multiplexing
   2. Dendritic backprop (spatial multiplexing)
   2. Feedback alignment
   3. Target propagation
   4. Relaxation methods - two phases
   1. Predictive coding
   2. Equilibrium propagation
   3. Prospective configuration
   5. Least-control principle (using the dual loss function) - single phase
   1. Feedback control
   2. Disinhibitory control