@vclay explores a new approach to symmetry in Monty, where Monty detects symmetry as it learns an object, storing that information directly in its model. Monty can then avoid initializing redundant hypotheses in future episodes, rather than rediscovering the same symmetry each time.
@vclay shares early prototyping results and how this idea is promising but tricky to implement. She explores the remaining questions, including how to recover when an object initially appears symmetrical but later reveals a distinguishing feature. The discussion closes with a broader look at mirror symmetry, including how the human brain learns to distinguish mirrored forms.
Summary Video
Main Video
0:00 Ideas on Representing Symmetry in Monty
0:08 Introductory Remarks
1:12 The Problem of Symmetry: What Monty Does Today?
3:20 Redundant Hypotheses Due to Symmetry
5:02 What are the Issues with Our Current Implementation?
6:31 Proposed Solution and Discussion: Store Detected Symmetry in the Object Models and Don’t Initialize Symmetric Hypotheses
14:12 Mug Example: Partial Symmetry
15:00 How Symmetry Marking Would Work in Practice
17:16 Expected Outcomes
22:08 Some Visualizations & Early Prototyping Results: What Current Detected Symmetry Looks Like
24:34 Iteratively Marking Off Symmetric Locations & Initializing From Them
27:07 Overall Performance Over 10 Episodes (Gradually Marking Symmetry)
27:31 Why are there no Reductions in the number of Steps or Runtime?
30:54 Tweaking Parameters
33:28 Open Questions and Issues
45:37 How Many Hypotheses Per Location?
56:15 Neural-Like Approach to Symmetry vs Abstract Symmetry Groups
59:36 Mirror Symmetry Question
1:00:04 AirPods: Left vs Right Asymmetry
1:01:47 Detecting Vertical vs Horizontal Symmetry in Humans
1:02:36 More Open-Ended Thoughts on Symmetry
1:05:32 Mirror vs Line Symmetry Definitions
1:12:11 Mirror Symmetry and Dyslexia
1:13:35 Mirrors are Hard to Break Paper: the Mirror Generalization Process
1:14:29 Further Discussions
1:22:11 Discussion on 3D Perception Bias and 2D Objects