@vclay reviews the different kinds of reference frames we have in Monty and how we transform between them. She covers the two scenarios where we need to compensate for orientation changes (role of the Thalamus): 1) When the sensor moves relative to the object and 2) when the object moves in the world. Some animations guide through these examples and explain how these transforms are achieved in Monty. For a write-up, see: https://thousandbrainsproject.readme…
Next, @nleadholm walks through an alternate proposal for voting on locations by voting on object centroids. Finally, the team discusses various questions and issues around voting and well as other potential solutions to the location sharing problem.
0:00 Introduction
2:29 Different Types of Reference Frames in Monty
20:45 Reference Frames Inside the Learning Module
22:03 Example: Object Rotates in the World
23:20 Example: Sensor Moving Relative to the Object
25:47 Summarizing the Two Mechanisms for Compensating Orientation Changes
31:57 Reference Frame Transformations During Voting
43:18 Two Mechanisms for Sharing Locations (Fixed Sensor Arrangement)
46:23 Alternative Idea: Voting on a Common Location
46:49 Creating a Global Location Representation
50:59 Example of Two LMs Sensing a “T”
52:15 Centroids Do Not Align for “+” (No Bag-of-Features)
53:44 Clarifications
54:44 Benefits
58:27 How Can Voting Happen if the Centroid Representation Changes?
1:00:04 Universal Encoding of Locations with Phases or Oscillations
1:01:32 Other Issues and Discussion
1:21:07 Clarifying the Location Sharing Problem During Voting
1:27:13 Can We Assume Fixed Sensors in Some Cases?
1:32:37 Can We Use Hierarchy for Knowledge Transfer Between Columns?
1:51:20 Walking Through a Simple Example of Voting