Greg Detre
Wednesday, May 07, 2003
Gallistel, �The principle of adaptive specialisation as it applies to learning and memory�, in Kluwe, Luer and Roesler, �Principles of human learning and memory�
adaptive specialisation in the memory mechanism
thermodynamic stability
high density
information is information
coding
adaptive specialisation in the learning mechanism
learning current location
learning the solar emphemeris
language
conditioning
Question to think about: Is curiosity a "thing", that� is, an adaptive specialization, or a collection of mechanisms that together produce what we call curiosity?
exploration vs exploitation tradeoff
temperature/stochastic
it�s the filter that tells you what concepts are interesting, or how to categorise your experience
learning about your environment
looking for a bunch of good cases/exemplars on which to rest your understanding, and which seem to be representative of the space � too specific to CBR
a methodology??? a drive??? � see lecture 030205
it�s got to be a set of qualities � what would a curiosity module look like???
well, unless maybe you�re in a safe niche and don�t need to be curious, except every so often
in which case it could maybe be something that gets turned on/off at need, kind of like a high-level model in a behaviour-based system (eating, sleeping, playing, followingMother, beingCurious etc.)
it could just be a parameter that affects the rest of the system
a means of evaluating how happy you are with current knowledge, i.e. whether you can survive on it if things stay the same, whether it�s barely adequate
a means of testing how much more there is to learn, or how fast the state of the world is changing, or whether you�re in a reasonably global optimum
a system for changing your current cognitive mechanisms/behaviour correspondingly, either by adding randomness, a bent for exploration, the interactions between them etc.
Emily: using previous knowledge to gain insight into new information � that�s just learning
me: exploit vs explore
Hugo: suitcase of heuristic mechanisms for principled state-space
discovery + learning
Andrea: actively seeking to understand/explain new experiences in the world
Derek: proactively exploring one�s environment in order to explore new strategies to maximise fitness � related to desire to play
Chris: behaviour that resolves uncertainties re novel circumstances, objects + phenomena
ability/desire to explore one�s space and extract features that enhance one�s understanding/functioning
desire to know something
Daphna: eager to learn more/ active desire to learn/know, interested in people + things around you
being proactive to learn about objects around you
Jessie: opportunistically seeking out information not required for current task
Cynthia: motivation to learn what it does not already understand and ability to act successfully on it
Deb: hunger [i.e. a basic disposition] for knowledge that
matters [of (evolutionary) significance to the organism]
Bruce: proactive goal-directed exploration with potentially distal integration and consequences
basically, it�s all just about going out and learning important stuff
being curious:
learning component
exploration
what �matters�, i.e. relevant to
you eventually
interesting/novel/unexpected
drive/proactive
knowledge may be most relevant
down the road
Deb: �curiosity� is just the drive
Derek: hunger for knowledge that doesn�t appear to matter
me: it�s where there may be a proximate reward if you learn something, but failing to learn won�t have a proximate cost (�you wouldn�t say you�re curious about how to get your leg out of a beartrap�) (if you fail to learn it)
Deb: C S Peirce � hard vs soft, things that sense and brace yourself vs compliance properties of rocks and sponge, things that one can theorise about that will never have consequence (and can�t be verified)
Deb: what matters is knowledge that you can make use of in servicing your highest level goals � benefit
Deb: it doesn�t make sense to talk of learning new goals at the highest levels
Jessie: not a module, but there is an innate bias towards things that are interesting/matter evolutionarily
Daphna: not a module
Kai-Yuh: neurologically encoded
e.g. music, not evolutionarily significant
brainwide neural activity
Deb: i.e. it�s implemented
general pattern that might let you generate new types of curiosity
Chris: boils down to implementation, multiple types of curiosity
Derek: neither, but especially not a unitary drive
consequence of not functional, metamorphic, play
Andrea: not a module, emergent quality of underlying (active) learning mechanisms
Hugo: same
critical mass of diversity of interests and competing mechanisms
me: affective set of parameters
weak method (in the Norman sense)
Deb: neither, see question 3
Cynthia: Andrea + Hugo camp - situatedness
Kai-Yuh: brain, vat (bat), machine � and I want a beach
related to past experiences
Deb:
analysis of ambiguity � reflect on what you don�t know
planning mechanisms to reduce ambiguity � e.g. changing your perspective or asking a question to resolve the ambiguity
constrained causal analysis of observations (analyses of causality) � trying to understand why what happened happened
Hugo:
three steps:
have a rough model
predictive forward projection
focus on some aspect � agenda or attention mechanism
situated model refinement � update that model
expectation-violation is a
trigger of curiosity
Deb: many types of curiosity, many types of triggers
Daphna:
whole cognitive architecture
ways to act on the world
beliefs + representations
Derek: mired too much in high-level cognitive representations � what about simple animals like cats
the real challenge is how to integrate results of opportunistic learning into continuing evolution of behaviour system � making best use of what you�ve discovered
Bruce: beavers � stick the stick into mud where the current is fastest
Deb: Griffin gave a counter-example of removing sticks sometimes from high-pressure areas
Jessie:
acquire + store information
drive to do that for seemingly unnecessary information
evolutionary bias towards being interested in things that turn out to be important
�
way to extract features from the environment
me:
a means of evaluating how happy you are with current knowledge, i.e. whether you can survive on it if things stay the same, whether it�s barely adequate
a means of testing how much more there is to learn, or how fast the state of the world is changing, or whether you�re in a reasonably global optimum
a system for changing your current cognitive mechanisms/behaviour correspondingly, either by adding randomness, a bent for exploration, the interactions between them etc.
OR me:
sense of what�s salient influencing your seeking-behaviour
whether you�re feeling confident/optimistic/inventive, i.e. safe + worth doing
curiosity as a (micro-???)stage like development, pruning, sleeping
Bruce:
trigger for when to be curious
exploration process
implicit mechanism to direct it � domain-dependent
when to stop being curious
integration phase
Bruce: reading a book recommended by Coppinger called �Individual development and evolution�
10 minute talks
couple of questions after each one