It's hard to imagine higher-level cognitive faculty without some form of hierarchical information processing. ATB proposed that such a hierarchy corresponds to the hierarchy of objects in the real world. This might be a bit too speculative. Columns learn from prediction errors. They can predict raw sensory inputs; they can also predict signals by other columns produced from sensory inputs. Thus, learning can happen when there are raw sensory input prediction errors as well as when there are other column signal prediction errors. Learning in columns can easily be hierarchical - naturally, models (or knowledge) learned from them are hierarchical. That being said, there is no reason to believe that the hierarchy is a neat pyramid with clear-cut division between layers. Any column can learn from any other columns as long as their signals are useful. It's just that learning, and thus models, can happen orders away from raw input signals.