It's interesting that when asked about why we use hierarchical methods in RL, ML, and cogsci, people haven't talked about the circuit complexity and information

It's interesting that when asked about why we use hierarchical methods in RL, ML, and cogsci, people haven't talked about the circuit complexity and information-theoretic reasons for using hierarchies. IIRC, a "deeper" circuit can represent a given function with exponentially fewer units/gates than a "shallower" circuit, and hierarchical methods also narrow down only the predictive information in the (supervised) dataset, according to information-bottleneck methods.