“Compositional Models and the Fundamental Problem of Vision”? Hierarchical Models

• One of the hopes, and expectations, of hierarchical models is that they can represent complex structures in terms of compositions of elementary components – shared parts. • This should yield big gains in the complexity of representation and inference. • But how can we analyze and quantify this? A Fundamental Problem of Vision • Complexity: • Set of images is almost infinite (Kersten 1987). • No. of objects is big 30,000 (Biederman 1984). • But the human brain can detect objects and understand scenes within 150 msecs. • And we want computer vision systems to do the same. The Fundamental Problem • This lecture explores this fundamental problem from the perspective of compositional models. • Quantify the gains of part sharing and executive summary. (Recall objects have a hierarchical distributed representation). • (I): We analyze compositional models and show they can yield exponential gains in efficiency. •.(II) We perform a similar analysis for a novel parallel implementation of compositional models. • (III) Speculations about the Visual Cortex