Greg Detre
Thursday, November 07, 2002
guests: Eric Mueller, Phil Apply(sp???), Oliver Selfridge
Eric �
Thought Treasure
Phil:
brokering urban transportation, offer to give people rides etc. to reduce
traffic
Oliver �
Pandemonium, coined �agent�
did someone in MAS962 say that Thought Treasure had folded???
well, it clearly hasn�t
www.massport.com � shows plane/airport arrivals/departures
guidebook for people travelling to Paris
take photos of a route, and say directions, uses voice recognition + CS to produce the directions
extending Aria to accept text stories, adding images
text game based on OpenMind � dynamically generates world
it would be a good way of exploring/testing the Open Mind database
you could almost have people entering data by their actions
you could even query the next Open Mind information-enterer for the next step in the story
demo: hoping to find the intersection of a bunch of different attributes to create a profile from the raw data in Open Mind, so that they can recommend something to someone given their various attributes (and their intersections)
Henry: spiral contexts (the union is too big, and the intersection is too small � could we use CS to expand the intersection just enough to get something done)
doesn�t this suffer from the frame problem???
�you have to not only know what�s relevant, but not
compute on what�s irrelevant
for symbolic systems, there�s no way of figuring out
what�s relevant just from the structure of the statements
�frame problem� is a phrase from psychology (rather
than Minsky/McCarthy), meaning a frame of reference within which all of the
issues are contained�
TouchGraph � visualisation of semantic nets, tension/proximity between concepts
how do you compare/evaluate this visualisation???
maybe you need to contextualise the results � there are just too many
actually, it could be pretty cool if you try and match up their profiles with an Amazon or IMDB database to generate recommendations
is there any CS involved here???
looking at cultural context in Open Mind
retrieving culturally-based information, and seeing conflicts
is there enough cultural diversity in terms of the inputters and what they�ve provided though???
how on earth do you do this??? how can you separate out cultural (rather than general/logical) conflicts???
how do you figure out whether you�ve got a conflict, or an exception, or no single default etc.???
Push: they don�t have much knowledge about things that can�t exist at the same time etc.
things can conflict with themselves � e.g. the barber who shaves everybody who doesn�t shave himself
Cyc does an initial shallow check for consistency, then does a longer chaining (sleep state/dreaming) search to see if contradictions pop up
interactive web browsing � learn background knowledge drawn from Open Mind about that topic
looking at WBBI platform
FlySwat generates links for almost every word linked into its internal database
use a virtual world (e.g. the Sims) to garner CS knowledge
what sort of CS knowledge might you get from Star Wars communities
network gaming e.g. Quake
knowledge-attacking the problem of common sense
you can mine the semantic nets (concepts connected by a set of predicates) in OM
no explicit context mechanism (although that�s implicit in how the graph is organised, and in the stories that prompted a given fact to be inputted)
keeps stories contiguously (amongst other representations)
http://www.research.ibm.com/people/e/etm
it seems to
me that Thought Treasure just won't ever be smart on its own
but it could be a pretty good scaffolding, like an automaton teacher,
for a slow-but-powerful learner system that might stand a chance of one day
being intelligent
to put it another way, it seems like it might be good as a skeletal
virtual environment in which a slow-learning-system could gambol
but Brooks would say that you risk defining your representations in
simple terms based on the wrong defining characteristics, that wouldn�t then
scale to the real world, and that the only way to form veridical/scalable
representations is to base/form them in the real world
no, he explicitly bills it as a tool for NLP etc.
TT has 500,000 assertions according to the CycL metric
much larger knowledge if you express it in CycL terms
because every word has assertions saying that this is a word etc.
Cyc has 1.5m assertions according to the CycL metric
Open Mind might then be bigger
why did he choose to make it multilingual??? why French???
I think it�s because he�s just got a thing about French, and this is all an enormous work of vanity anyway
manually entered all the knowledge himself
semantic component vs understanding agency???
the understanding agency takes all the low-level stuff (including the predicate logic representation) and tries to build a much more detailed model
to understand stories, model the states + events in the story
modularise: different agents work on different parts of the story
lucid � where everything is very concrete (i.e. unified???), vs indirect (e.g. propositional knowledge)
grids � ascii art
but how do you deal with fuzziness, scaling etc. with the grids???
how do you deal with viewing the same scene from different perspectives???
the wormholes are too restrictive as a joining mechanism
can two things occupy the same space???
no, that�s why he has shortcut/filler symbols to show platform, where people can be
the ascii grids get parsed out and stored as assertions
there�s now a directory of the whole of Paris, showing what shops are on which street etc.
the problem with the idea of multiple representations is that you just shift the problem to how to relate those representations
he seems to deal with that by giving probabilistic howMuchSenseThisMakes ratings that can be compared across representations
how does it deal with unexpected problems???
having all the timeouts everywhere (e.g. in the PhoneCall planning agent) just adds more detail than necessary � he can�t be imagining that this would actually scale to real world behaviour
having said that, there is also a bit that simulates what the phone actually does � presumably you could make that real-world complex easily � but it wouldn�t feed back/teach the PhoneCall planning agent to deal with it in more correspondingly complex ways
it has a whole �hello� ontology
rich but brittle
using multiple representations is only as useful as your means of getting them to work together
also, I don�t see how it could garner new information for itself � how does he think he can (semi-) automate it???
he�s hoping to use Open Mind etc.
SensiCal � smart calendar (�you�ve booked lunch with a vegetarian in a steakhouse�)
Hugo: the killer app for CS is NLP
worth looking at the CS of tasks, priorities etc.
of course, the CS of calendaring is AI-complete
what does he mean by the �hard� common sense problem???
refce: Henry: interesting paper, serious study on �How near is near?� (Murray Denovsky(sp???))
Watts
and Burgess (1981) � imagery, stick figures
if you have a spreadsheet that you want to do a sanity check of, then you have an ontology-matching problem (matching the �age� column to your age representation)
Cyc�s synonym matching is too sparse (based on Wordnet)
open mind:
how do participants design the templates???
how are the templates organised/related to each other???
is the information from the first generation system still available/useable???
does an MIT human read/verify everything that comes in???
does it find paradoxes (cf cheap apartments) a lot???
how does it place things in context???
does it generate its own questions???