This is in response to the deep predictive learning thread. I present an alternative hypothesis, which also has predictive learning, error signals, an auto-encoder, and it is (to my knowledge) biologically constrained. However I have not diligently researched this idea, nor written very many pages on it.
First I recommend reading “Thalamic relays and cortical functioning” by S. Murray Sherman, DOI: 10.1016/S0079-6123(05)49009-3 . Figure 4 is especially relevent.
Conjecture: The Thalamic Reticular Nucleus (TRN) decodes cortical activity into the sensory inputs which likely generated it, and inhibits those inputs.
Explanation: The cortex builds a predictive model of the world. If the cortical model is working correctly then it should be able to account for all of its sensory inputs. The errors/discrepencies in this accounting are a powerful learning signal. This is the idea behind auto-encoders.
Cortical layer 6 contain a location signal. The locations are unique to the object/reference-frame they are on. This is exactly the information which is needed to predict the current sensory input. L6 projects back to the TRN, which is GABAergic and in turn inhbits the Thalamic Relay Cells. The TRN selectively inhibits the sensory inputs which the cortex understands, so that the cortex can focus on the unexplained inputs.