Michael Hodel: Reverse Engineering the Abstraction and Reasoning Corpus

Had a great discussion with Micahel Hodel and a few others (Simon Strandgaard, Yassine and many more) about reverse engineering the ARC dataset, and possible approaches to solve ARC.
Speaker Profile:
Michael studied Computer Science in Zurich, Switzerland. He is currently working as a freelance programmer, but his main focus remains ARC. Currently, his team MindsAI with Jack Cole and Mohamed Osman is #1 on the ARC-Prize leaderboard (a 1 million dollar competition to solve ARC - arcprize.org/ )
Abstract:
The Abstraction and Reasoning Corpus (ARC) is a dataset intended to serve as a benchmark for general intelligence. The difficulty of ARC for machine learning approaches is largely a consequence of the great diversity of tasks as well its few-shot nature. Even after almost five years since its publication, ARC remains unsolved. While many attempts have been made to solve the benchmark, what seems generally lacking are more fundamental scientific experiments. RE-ARC presents code to procedurally generate examples for the ARC training tasks and with that attempts to enable experiments addressing the latter of those two aspects, namely sample-efficient learning. RE-ARC also introduces a simple proxy metric for example difficulty, which should also allow for exploring questions about within-task generalization capabilities of systems.
Repo: github.com/michaelhodel/re-arc
Paper: arxiv.org/abs/2404.07353
~~~
0:00 Speaker Introduction
0:45 Introduction to ARC-DSL
8:26 Data Generation
12:13 How close is DSL to human priors
14:42 How to decide which DSL to keep and which to add in
19:53 Introduction to RE-ARC
23:45 Overview of RE-ARC
25:32 Task Generalisation in RE-ARC
26:00 Example Verification in RE-ARC
26:47 Example Difficulty in RE-ARC
31:48 Limitations of RE-ARC
33:37 Examples of RE-ARC
35:22 Using RE-ARC to gauge model learning
37:48 Vision for meta-learning beyond RE-ARC
39:08 Can arbitrary DSL be generated with RE-ARC?
43:17 Discussion
~~~
AI and ML enthusiast. Likes to think about the essences behind breakthroughs of AI and explain it in a simple and relatable way. Also, I am an avid game creator.
Discord: / discord
LinkedIn: / chong-min-tan-94652288
Online AI blog: delvingintotech.wordpress.com/
Twitter: / johntanchongmin
Try out my games here: simmer.io/@chongmin

[@johntanchongmin]

0:00 Speaker Introduction 0:45 Introduction to ARC-DSL 8:26 Data Generation 12:13 How close is DSL to human priors 14:42 How to decide which DSL to keep and which to add in 19:53 Introduction to RE-ARC 23:45 Overview of RE-ARC 25:32 Task Generalisation in RE-ARC 26:00 Example Verification in RE-ARC 26:47 Example Difficulty in RE-ARC 31:48 Limitations of RE-ARC 33:37 Examples of RE-ARC 35:22 Using RE-ARC to gauge model learning 37:48 Vision for meta-learning beyond RE-ARC 39:08 Can arbitrary DSL be generated with RE-ARC? 43:17 Discussion

[@oncedidactic]

great to have this interview!

[@johntanchongmin]

Repo: github.com/michaelhodel/re-arc Paper: arxiv.org/abs/2404.07353