In this research meeting, we discussed top-down connections and other topics related to modeling compositional objects. @nleadholm presented some potential problems with sparse compositional models and ambiguity that could arise. We talk through those problems and potential solutions and conclude that they are not fundamental issues. We also touched on concepts like attention and leveraging different model types across columns.
Summary Video
Main Video
0:00 Introduction
0:58 Top-Down Projections: Logo on a Mug
3:10 A Problem with Sparse Models
8:57 Potential Ways of Dealing with this Ambiguity
20:13 Could we just Predict a Union?
29:06 More than 2 LMs: Which LMs Would Represent What?
37:16 Leveraging Voting
39:44 Attending to Regions in Space
47:32 Further Discussions: More General Sensor Modules
54:34 Teaser: Exciting Work in Progress!