On the Measure of Intelligence

by François Chollet

To make deliberate progress towards more intelligent and more human-like artificial systems, we need to be following an appropriate feedback signal: we need to be able to define and evaluate intelligence in a way that enables comparisons between two systems, as well as comparisons with humans. Over the past hundred years, there has been an abundance of attempts to define and measure intelligence, across both the fields of psychology and AI. We summarize and critically assess these definitions and evaluation approaches, while making apparent the two historical conceptions of intelligence that have implicitly guided them. We note that in practice, the contemporary AI community still gravitates towards benchmarking intelligence by comparing the skill exhibited by AIs and humans at specific tasks such as board games and video games. We argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to "buy" arbitrary levels of skills for a system, in a way that masks the system's own generalization power. We then articulate a new formal definition of intelligence based on Algorithmic Information Theory, describing intelligence as skill-acquisition efficiency and highlighting the concepts of scope, generalization difficulty, priors, and experience. Using this definition, we propose a set of guidelines for what a general AI benchmark should look like. Finally, we present a benchmark closely following these guidelines, the Abstraction and Reasoning Corpus (ARC), built upon an explicit set of priors designed to be as close as possible to innate human priors. We argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans.

Genres: Artificial Intelligence, Nonfiction, Technology

ebook

Published November 1, 2019

About the author

François Chollet is a French engineer and researcher in artificial intelligence.

Reviews

[Rahim Hashim]

If you're reading this review, you're familiar with François Chollet and his work. If you’re familiar with his work, you already know that you should read anything and everything he writes.

This is certainly not an exception—François does an incredible job of breaking down the incredibly nuanced problem of defining intelligence in such a way that experienced ML researchers and young students of the field alike can appreciate. It goes without saying that his work broadly, and this paper specifically, will impact the ways in which we think about intelligent systems—as optimization problems, where the objective function is a computable approximation of intelligence. Framing intelligence as a measure of skill-acquisition efficiency, and then providing a benchmark (Abstraction and Reasoning Corpus, ARC), makes the problem less abstract and more concrete, which is critical for those thinking about how to develop the systems.

Studying behavioral neuroscience, I cannot recommend this paper more for others doing the same, and hope that many in other fields/industries will take the time it requires (no easy feat, given certain sections are packed with theory and formalizing terms) to read and understand it.

Contextualizing this review, as I'm writing this in Oct 2020, it feels like liberal democracy is crumbling globally before our very eyes. It may require researchers, scientists, teachers, students, and citizens together to progress AI in a significant in order to save us from a very dark timeline. I hope this changes the direction of our society's objective function, for the better.

[taylor]

This paper is still very much relevant after 5 years. I will reread several times and hopefully have a firmer understanding of the idea of "intelligence". Perhaps I can even attempt to build something intelligent in a small environment. My synthesis of the paper is that intelligence is the ability to learn and adapt to new situations, where new is a function of your prior experience. I do not believe we become more intelligent just adding hierarchically to the catalog of priors. Learning to a large degree is a prior, innate attribute. Intelligence must include active learning. Bravo, what a great read.

[Davidius]

Intelligence as skill-acquisition efficiency.

AI as the science & engineering of making machines do tasks they've never seen & have not been prepared for.

Both of these views of the nature of human intelligence – either a collection of special-purpose programs or a general-purpose Tabula Rasa – are likely incorrect.

If intelligence lies in the process of acquiring skills, then there is no task X such that skill at X demonstrates intelligence, unless X is actually a meta-task involving skill-acquisition across a broad range of tasks.

A fundamental notion in psychometrics is that intelligence tests evaluate broad cognitive abilities as opposed to task-specific skills.

most benchmarks do not pay much attention to formally assessing robustness and quantifying generalization, and thus can be solved via “shortcuts” that gradient descent is apt at exploiting

the hallmark of broad abilities (including general intelligence, as per II.1.2) is the power to adapt to change, acquire skills, and solve previously unseen problems

4 areas of Core Knowledge encoded by evolution, reflecting long-term stable environmental aspects associated with evolutionary pressure:
- objectness & elementary physics
- agentness & goal-directedness
- natural numbers & elementary arithmetics
- elementary geometry & topology

The intelligence of a system is a measure of its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience, and generalization difficulty.

generalization describes the capability to deal with future uncertainty, not the capability to compress the behavior that would have been optimal in the past – being prepared for future uncertainty has a cost, which is antagonistic to policy compression
- education of children involves practicing games and ingesting knowledge of seemingly no relevance to their past or present decision-making needs, but which prepare them for future situations (a process often driven by curiosity
- implies the need to store representations of past data that would be seemingly useless from the perspective of the past but may prove useful in the future

intelligence is the rate at which a learner turns its experience and priors into new skills at valuable tasks that involve uncertainty and adaptation.

Consequences for research directions towards flexible or general AI:
- encourages emphasis on program synthesis as part of the system that possesses intelligence (“intelligent system”, a program-synthesis engine) from the part that achieves skill or implements behavior (“skill program”, the non-intelligent output artifact of the process of intelligence)

generate tasks via an ever-learning “teacher” program, interacting in a loop with test-taking “student” programs
- The teacher program would optimize task generation for novelty and interestingness for a given student (tasks should be new and challenging, while still being solvable by the student)
- students would evolve to learn to solve increasingly difficult tasks
- favors curriculum optimization, as the teacher program may be configured to seek to optimize the learning efficiency of its students...In order to make sure that the space of generated tasks retains sufficient complexity and novelty over time, the teacher program should draw information from an external source (assumed to feature incompressible complexity), such as the real world