@sknudstrup discusses a new attention system prototype for Monty that combines bottom-up salience with top-down signals from learning modules. The short-term goal is to constrain movement and sensory input to a specific object or sub-object until it has been recognized, then release attention so Monty can move on to recognize the next object.
He walks through the current prototype, and discusses 3D attentional regions, voxel-based representations, excitatory and inhibitory signals, and hard versus soft filtering of motor goals. The conversation also includes open questions around scale, moving objects, inhibition of return, and how unexpected events might interrupt attention.
Summary Video
Main Video
0:00 Introduction
0:08 Fate Attenzione!
0:56 Primary Needs for an Attention System
4:38 SM Attentional Regions
6:18 Discussion: Attention Scale and Zooming In and Out
18:01 LM Attentional Regions
19:30 Discussion: How to Generate Point Clouds from Learning Modules
28:48 Discussion: Accumulating Locations and Releasing them in the Attention System
34:03 When to Give Up Attention on Achieving a Goal? The Puzzle Example
38:12 Initiating Voxels with a Weight Implementation and Discussion
46:50 Discussion: A Reset Signal for Attention
51:33 Discussion: Voting & the CMP for Attention Signals
56:11 Discussion on Scale for Attention & Scale Terminology: Model-Based vs Top-Down
1:05:39 Discussion: What Would Make Scale Parameters Meaningful?
1:13:15 Modeling Monty’s Position and Perspective
1:22:11 Off-Object Observations and Hypothesis Elimination
1:24:30 Wrap-Up and Next Steps