@vclay and @nleadholm discuss attentional regions, policies, and prediction error. They wonder about how to have Monty focus its attention on areas in space instead of discrete points. They discuss how attention relates to policies, gating sensory input, and voting. Specifically, the team explores how they could implement policies to stay on an object until it’s recognized to help with object recognition. There are high-level questions about how model-free policies are defined in the first place, and whether prediction error (both in the long and short term) could be used as a general framing.
Summary Video
Main Video
0:00 Introduction
1:35 Target Pose, Attention Area and how They Relate to Monty’s Policies
7:44 Adding Attended Regions and Gating
23:42 Are Attentional Areas Defined by a Set of Cortical Columns?
34:37 Are Attentional Areas Amorphous Followed by Narrowing Down Location?
40:03 “Stay on Object until Recognized” Policies
52:29 Could Modeling with Multiple Columns Instead of One Make The Problem Easier to Solve?
1:12:42 Using Model-Based Information to Define Attentional Regions
1:20:05 Voting and Anchoring Grid Cells
1:37:20 The Hypothesis Testing Policy and Other Model-Based Policies
1:48:48 Further Discussions