Part of my series of notes from ICLR 2019 in New Orleans.
There are a lot of interesting papers referred to in this talk, rather than link to them individually I’ll point you to Oudeyer’s website, go wild.
Introduction
- human babies are crazy and learn all kinds of shit on their own
- body morophology – emergent / self-organized structures & motions
- cognitive biases – e.g. for language acquisition (affordances)
- intrinsic motivation – curiosity, self-motivated exploration, active learning
Curiosity-driven exploration
- benchmarks on intelligence can be very misleading
- “what the hell, the point was to get the toy out of the tube!”
- related to theory of flow
- (a good read if you’re interested…)
- want real theoretical framework in psych / neuroscience
- investigate using robotic playgrounds
- movement and perceptual (object-based) primitives – higher level than pixel-level or micro-muscle movements
- many ideas of what makes an interesting learning experiment…
-
learning progress hypothesis
- interestingness proportional to empirical learning progress (absolute value of derivative)
- how to do this without the global view?
- hierarchical multi-armed bandits – partition space based on learning difficulty
- inverse model (goal) exploration vs. forward model exploration
- robotics spaces have lots of redundancies – exploring many ways to get small number of effects, vs. fewer ways to get large number of effects
- discovers nested tool use (i.e. using tools that control other tools…)
- MUGL: exploring learned modular goal spaces
- learning from pixel space
- learning beta-VAE representation spaces for exploring goals
Back to humans…
- modeling child development data
- e.g. vocal development
- emergent developmental stages – no sounds, unarticulated sounds, articulated sounds
- shift from self-exploration to imitation (this is cool)
- individual differences in developmental trajectories – different attractors
- discover language as a tool to manipulate environment
-
The Ergo-Robots! :robot_face:
- self-organization of culturally-shared speech sounds
- bias to sync up noises with others
- similar vowel families developed as for human languages (!)
Image above included solely for my phonology friends, who might even understand it.
- interplay of regularity & diversity
- applications to educational technologies
- personalize curriculum for efficient learning and intrinsic motivation
Q&A
- can develop communication without having a model of other minds, whoa
- how to do this without human-provided representations and similarity metrics for goal space?
- not done for robot experiments (for now), happens in beta-VAE experiments though
- some additional work on metric learning for embodied systems that he didn’t talk about…