More-or-less the same principle, but scaled up massively, and with context-dependent probability conditioning maps.