@vclay reviewed last week’s proposal to remove state from the morphology models. She went over the main problem we are trying to solve, the way that morphology models with state differ from those without, and the open questions and issues that arise from that. She also proposed a potential alternative solution to the initial problem that would keep the state in morphology columns by utilizing unions of them to apply different states to different locations on an object.
The team discusses various issues and questions around those topics and also brainstorms around broader, different mechanisms, such as learning the Gestalt of something without utilizing reference frames.
Summary Video
Main Video
0:00 Full
2:47 The Problem
6:36 Removing State from Morphology Models - Comparison, Implications, and Issues
22:09 What if We Could Just Solve the Initial Problem in a Different Way?
1:03:12 Example: Combination Lock
1:13:52 Issues with State Applying to Parts of the Object
1:20:23 Further Discussion & Questions
1:46:16 Is There Another Mechanism at Play? Gestalt-like Representations?
2:17:50 Paper: Object and Feature Representations in Macaque Monkey