From OpenCog
Jump to: navigation, search

The Perceptual-Motor Hierarchy

The perceptual approach outlined in OpenCogPrime:PerceptualPatternMining is "flat," in the sense that it simply proposes to recognize patterns in a stream of perceptions, without imposing any kind of explicitly hierarchical structure on the pattern recognition process or the memory of perceptual patterns. This is different from how the human visual system works, with its clear hierarchical structure, and also different from many contemporary vision architectures, such as Hawkins' Numenta system which utilizes hierarchical neural networks.

However, the approach described above may be easily made hierarchical within the OCP architecture, and this is likely the most effective way to deal with complex visual scenes. Most simply, in this approach, a hierarchy may be constructed corresponding to different spatial regions, within the visual field. The RegionNodes at the lowest level of the hierarchy correspond to small spatial regions, the ones at the next level up correspond to slightly larger spatial regions, and so forth. Each RegionNode also correspond to a certain interval of time, and there may be different RegionNodes corresponding to the same spatial region but with different time-durations attached to them. RegionNodes may correspond to overlapping rather than disjoint regions.

Within each region mapped by a RegionNode, then, perceptual pattern mining as defined in the previous section may occur. The patterns recognized in a region are linked to the corresponding RegionNode — and are then fed as inputs to the RegionNodes corresponding to larger, encompassing regions; and as suggestions-to-guide-pattern-recognition to nearby RegionNodes on the same level. This architecture involves the fundamental hierarchical structure/dynamic observed in the human visual cortex. Thus, the hierarchy incurs a dynamic of patterns-within-patterns-within-patterns, and the heterarchy incurs a dynamic of patterns-spawning-similar-patterns.

Also, patterns found in a RegionNode should be used to bias the pattern-search in the RegionNodes corresponding to smaller, contained regions: for instance, if many of the sub-regions corresponding to a certain region have revealed parts of a face, then the pattern-mining processes in the remaining sub-regions may be instructed to look for other face-parts.

This architecture permits the hierarchical dynamics utilized in standard hierarchical vision models, such as Jeff Hawkins' and other neural net models, but within the context of OCP's pattern-mining approach to perception. It is a good example of the flexibility intrinsic to the OCP architecture.

Finally, why have we called it a perceptual-motor hierarchy above? This is because, due to the embedding of the perceptual hierarchy in OCP's general Atom-network, the percepts in a certain region will automatically be linked to actions occurring in that region. So, there may be some perception-cognition-action interplay specific to a region, occurring in parallel with the dynamics in the hierarchy of multiple regions. Clearly this mirrors some of the complex dynamics occurring in the human brain.