This invited talk on 'Creativity is Intelligence' was given by Kenneth Stanley at the 2018 NeurIPS Workshop on Machine Learning for Creativity and Design in Montreal, Canada on the 8th December.
Creativity is Intelligence - Kenneth Stanley - NeurIPS Creativity Workshop
1. Creativity Is Intelligence
Kenneth O. Stanley
Uber AI Labs
And
Evolutionary Complexity Research Group,
Department of Computer Science,
University of Central Florida
kennethostanley@gmail.com
kstanley@uber.com
2. To Solve Is to Create
• Some problems are solved by following the gradient of
performance
3. To Solve Is to Create
• Some problems are solved by following the gradient of
performance
– But these are the easy ones
4. To Solve Is to Create
• Some problems are solved by following the gradient of
performance
– But these are the easy ones
• For the rest, we must seek the novel and the interesting
– The foundation of creativity
18. What We Discovered: People Only Find
When They Are Not Seeking
The stepping stones almost never resemble the final product!
Moral: You can only find things by not looking for them
Stepping stone to the Teapot
Stepping stone to the Skull
Stepping stone to Jupiter
Stepping stone to the Penguin
Stepping stone to the Butterfly
Stepping stone to the Lamp
20. Leads to New Class of Automated Algorithms
Starting with Novelty Search (NS)
• Originally from Lehman & Stanley 2008, 2011
– Raised idea of divergent search algorithms without specific objectives
– Divergent search instead of convergent search
• Intriguing early results show NS sometimes solves better than optimization
21. NS Leads to
Quality Diversity (QD) Algorithms
• Combine novelty with a quality objective
– Spawns a new field: novelty search with local competition (NSLC), Multi-
dimensional Archive of Phenotypic Elites (MAP-Elites), etc.
– Find the best possible representatives of diverse niches: a repertoire
• (so far not contemplated in RL)
Cully, A., Clune, J., Tarapore,
D., and Mouret, J.-B. "Robots
that can adapt like animals."
Nature, 521.7553 (2015)
Lehman, Joel, and Kenneth
O. Stanley. "Evolving a
diversity of virtual creatures
through novelty search and
local competition."
Proceedings of the 13th
annual conference on
Genetic and evolutionary
computation. ACM, 2011.
26. Fun with Divergence: Galactic Arms Race
• Game: invents its own weapons as the game is
played
Ultra-Wide CorkScrew Ladder
Wall Gun Tunnel Maker Trident
(Hastings, Guha, and Stanley 2009)
27. Fun with Divergence: Petalz
Risi, S., Lehman, J., D'Ambrosio,
D., Hall, R., & Stanley. Petalz:
Search-based procedural content
generation for the casual gamer.
IEEE Transactions on
Computational Intelligence and AI in
Games, 2015.
28. But Remember…
To Solve Is to Create
• Quality Diversity means collecting diverse stepping
stones in an archive
– A form of creativity that solves problems
• What if instead of collecting interesting policies of the
past we collected interesting states?
– QD of state space
29. Answer: We Get Go-Explore
(just announced from Uber AI Labs)
with Adrien Ecoffet, Joost Huizinga, Joel Lehman, Jeff Clune
• Try to get somewhere novel and return someday to see if
you can find something even more novel (diversity)
• But if you find a better path to the same place again,
replace it in the archive (quality)
• An exploration of states inspired by open-endedness,
QD, and MAP-Elites
• So what happens?
(e.g. backwards
learning
from Salimans and
Chen 2018)
30. Answer: We Get Go-Explore
(just announced from Uber AI Labs)
Adrien Ecoffet, Joost Huizinga, Joel Lehman, Ken Stanley, Jeff Clune
https://eng.uber.com/go-explore/
32. And More Is Coming Soon:
POET (Paired Open-Ended Trailblazer)
• With Rui Wang, Joel Lehman, Jeff Clune
• Can we open-endedly invent new problems and optimize
solutions to those problems indefinitely?
– Borrows from Minimal Criterion Coevolution (Brant and
Stanley), CMOEA (Huizinga and Clune), and MAP-Elites
(Mouret and Clune)
• Inspiration: If you want to invent new solutions forever,
then you need to invent the problems too
– The solutions to old problems become stepping stones to new
ones
• Hypothesis: Only way some solutions can ever be found
35. And all Kinds of Innovative Skills,
All in the Same Run
“Downstairs Hero”
36. Creativity Stretches our Understanding of
Intelligence
• Opens a huge spectrum of possibilities because it is a
huge component of intelligence
• Even in less obviously creative endeavors
37. More information
• My Homepage: http://www.cs.ucf.edu/~kstanley
• Evolutionary Complexity Research Group:
http://eplex.cs.ucf.edu
• Uber AI Labs: https://www.uber.com/info/ailabs/
• Email: kennethostanley@gmail.com
kstanley@uber.com
• Twitter: @kenneth0Stanley
• https://eng.uber.com/go-explore/
zero
Go-Explore: