The team shares project ideas for the upcoming Rome Monty focus week hackathon.
Several proposed projects focus on unsupervised learning: how can Monty learn objects it hasn’t seen before, expand its object categories, and decide when to combine models or keep them separate. Other ideas include connecting Monty to the ARC-AGI-3 test bed, and using symmetry and attention to help Monty recognize and explore objects. There are some engineering project proposals too, like adding telemetry for data collection, and architectural improvements to how sensor and learning modules operate.
Main Video
0:00 Introduction
0:07 Hackathon Introduction & Format
2:05 Project: Unsupervised Learning of Child Objects
8:59 Project: Explore Unsupervised Learning, Hierarchy & Categories
11:47 Project: Consolidate Architecture Design (One Architecture to Rule Them All)
17:23 Project: Reference Frame Merging through Inference
22:28 Summary of Unsupervised Learning Projects
29:16 Project: Monty Plays Gameboy (An ARC-AGI-3 Integration)
38:09 Project: It’s all Greek to me (Omniglot test bed)
43:45 Project: Turn Symmetry into a Strength Instead of a Weakness
44:46 Project: Attention, please! (Implement a Monty Attention System Prototype)
47:48 Project: Telemetry…telemetry everywhere…
49:38 Project: Sensor Module as a Pipeline of Transforms
52:52 Project: Learning Module as a Pipeline of Transforms
1:01:41 Project: Move motor_only_step to the Simulator
1:05:50 Voting Phase