Top-Down Influences on Bottom-Up Processing
Abstract
Perception is not simply a bottom-up process, but involves inductive inferences that use top-down knowledge to interpret image data. In computational vision, typically this knowledge appears in the form of constraints such as rigidity, viewpoint consistency, smoothness, etc. However these constraints are fallible - they do not always apply, and hence the perceptual process entails inductive reasoning. A major theoretical thrust of our work is to provide a formal lattice framework for organizing the plausible states of this reasoning aspect of perception (Jepson & Richards). The proposal makes strong predictions, given a set of constraints and a particular picture as to what interpretations or percepts will be seen. Consequently we have a series of experiments underway to (1) understand which constraints (or premises) are typically invoked in the interpretation of simple line drawings and (2) to show that the 'lattice framework specifies all of these interpretations, placing them in proper rank order. In parallel, we are also exploring two other models for merging bottom-up and top-down information, both of which are neural-based. One, called sequence- seeking (Ullman), proposes a network hierarchy where a sequence of transformations of both the input data and the target model occur in parallel, searching for the proper mapping that brings each into register. The proposal makes a special effort to incorporate what we currently know about cortical machinery, and also has triggered psychophysical experiments. (We have not yet explored the relations between the lattice and sequence-seeking proposals.) Finally, there are some studies related to our ability to switch between sets of premises, or to alter our models.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 04, 1992
- Accession Number
- ADA248479
Entities
People
- Whitman Richards
Organizations
- Massachusetts Institute of Technology