[Paper Review] AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence

Jeff Clune|arXiv (Cornell University)|May 27, 2019

Machine Learning and Data Classification189 references52 citations

TL;DR

The paper argues for AI-generating algorithms (AI-GAs) as an alternate path to general AI, outlining three pillars—meta-learning architectures, meta-learning learning algorithms, and generating effective learning environments—and discusses why AI-GAs may be the fastest route while still valuing the manual path.

ABSTRACT

Perhaps the most ambitious scientific quest in human history is the creation of general artificial intelligence, which roughly means AI that is as smart or smarter than humans. The dominant approach in the machine learning community is to attempt to discover each of the pieces required for intelligence, with the implicit assumption that some future group will complete the Herculean task of figuring out how to combine all of those pieces into a complex thinking machine. I call this the "manual AI approach". This paper describes another exciting path that ultimately may be more successful at producing general AI. It is based on the clear trend in machine learning that hand-designed solutions eventually are replaced by more effective, learned solutions. The idea is to create an AI-generating algorithm (AI-GA), which automatically learns how to produce general AI. Three Pillars are essential for the approach: (1) meta-learning architectures, (2) meta-learning the learning algorithms themselves, and (3) generating effective learning environments. I argue that either approach could produce general AI first, and both are scientifically worthwhile irrespective of which is the fastest path. Because both are promising, yet the ML community is currently committed to the manual approach, I argue that our community should increase its research investment in the AI-GA approach. To encourage such research, I describe promising work in each of the Three Pillars. I also discuss AI-GA-specific safety and ethical considerations. Because it it may be the fastest path to general AI and because it is inherently scientifically interesting to understand the conditions in which a simple algorithm can produce general AI (as happened on Earth where Darwinian evolution produced human intelligence), I argue that the pursuit of AI-GAs should be considered a new grand challenge of computer science research.

Motivation & Objective

• Present the manual AI approach and its challenges in assembling many building blocks for general AI.
• Introduce AI-GAs as a scalable alternative that learns how to build general AI without hand-designing every component.
• Outline the three pillars essential to AI-GAs and discuss expected research directions and safety considerations.
• Advocate for shifting research investment toward AI-GAs while acknowledging the value of the manual path.

Proposed method

• Describe the Three Pillars: (1) meta-learning architectures, (2) meta-learning the learning algorithms, (3) generating effective learning environments.
• Compare AI-GA to the manual approach and discuss how automated learning of architectures, learners, and environments could enable general AI.
• Review prior work and propose research directions for each pillar, including architectures search, learned optimizers, and curriculum/data generation.
• Argue for a grand-challenge framing of AI-GAs and discuss potential safety and ethical considerations.

Experimental results

Research questions

• RQ1 Can an AI-generating algorithm (AI-GA) bootstrap from scratch to produce general AI via three interconnected pillars?
• RQ2 What research directions and technical challenges exist for meta-learning architectures, meta-learning learning algorithms, and automatic environment/curriculum generation?
• RQ3 Under what conditions could AI-GAs outperform manual AI pathways in producing general intelligence?

Key findings

• Architecture search has begun to outperform hand-designed architectures on benchmarks like CIFAR and ImageNet, indicating potential in the First Pillar.
• Meta-learning approaches (e.g., MAML and RNN-based meta-learners) can yield faster adaptation and internally guided learning strategies, with methods like differentiable plasticity and neuromodulation improving continual learning.
• Generating effective learning environments and curricula is identified as the least explored but hardest pillar, with potential to meaningfully shape the learning process via task distributions and training data.
• The AI-GA framework could enable scalable, open-ended progress and provide insights into the origins and space of possible intelligences, while offering short-term value even if long-term goals are not fully realized.
• The paper discusses safety and ethical considerations unique to AI-GAs and advocates for treating AI-GAs as a grand scientific challenge and a potential fastest path to general AI.