Greg Detre
Thursday, November 14, 2002
Bill Woods (Sun) � �What�s in a link?� (semantic networks)
IBM � new Common Sense effort in Watson (NY) � Doug, L Morgenstern, Erik Mueller, logician, statistician
Minsky � causal diversity matrix???
NSA � Danny Hillis� company (Applied Minds) involved
they�re also trying to use Cyc
ARDA
contribute to the bulletin board re cheap apartments etc.
papers for next week
Whirl (William Cohen) � interleaving search + inference
OMCS toolkit API available
query the OMCSNET for a word/concept
connect two concepts
spreading activation � find all somewhat related concepts (can you do intersections with that???)
Phil left a stack of Cyc memos with Push
trade-off between work you do on the front end to put the knowledge in, and work you do on the back end to get the knowledge out usefully
pga@ai.mit.edu
not affiliated with Cycorp
MIT AI Lab, Music cognition group, UI, Selfridge Pandemonium feature recogniser for fonts
DEC Cambridge research lab � intelligent calendaring + Personal information Management
evaluate/critic Cyc, started building his own Knowledge Editing Tools and Methodology
DEC withdrew $2.5m for Cyc
two representation languages
epistemological level language
predicate calculus with reification (2nd order)
elegant + uniform
but you can�t compute in it because of the computational complexity and undecidability
heuristic level language
variety of representations for efficiency of expression + inferencing
> 20 inferencing mechanisms with specialised representations
general truth maintenance system
default reasoning (requiring you to assert all the exceptions)
context mechanism later (early 90s)
supposed to be able to automatically translate between E- and H-level
context
supposed to be lightweight mechanisms
lifting rules???
e.g. what�s considered affordable in different countries
most concepts have different interpretations in different contexts
e.g. in an American context, the concept of a person � an American
Henry: traditional logic doesn�t have any place to put �under these circumstances�� � how about quantifiers???
Lenat � dimensions of context space
Alan Kay � said of Lenat that he�s the master of taxonomising
taxonomies of context
problem solving context
event context
counterfactual context
hypothetical context
theory context
edit session context
consensus reality + context
fallacy of consensus reality and the granularity of context � too little too late
afterthought
not at proper level of granularity
didn�t go back and re-implement assertions within contexts
did not include information in context pertaining to applicability: appropriate problem-solving contexts, completion state, teesting and validation, use history
did not focus on discourse contexts or other fine-grained contexts
management of the KR process
KE-Region
bases new concepts on templates
copy/paste/edit approach
allows incomplete/inconsistent/questionable assertions to be recorded the user is ready to install them
associates commentary with ssertions
groups related assertions
doesn�t scale
documentation doesn�t get updated
insufficient feedback
situated learning
does CS require a body?
ted: in our life, things get easier the more knowledge you have
do things get more complex when they get more complex?
this is the thing about human-built complex systems � they�re never robust
build agencies out of context later (Lenat idea)
common lisp interface manager (CLIM) � now open source
visual future
object that refers to an incomplete computation state
continue your computation until you need that result
self-reflective tools for managing the knowledge entry process
he thinks that if the regimented procedure he�s proposing had been used from the start, the Cyc database might be in a much better situation
Henry: if you make the knowledge entry task so difficult, then you just might not get enough knowledge
how would his little procedural tips help with the quality of the knowledge being entereed?
his claim is that there were just little errors and omissions being made all over the place, and not being picked up
it�s basically just a rigorous testing and bug-tracking/code-management system � just good software-engineering habits
DEC�s main interest in Cyc was probably the calendaring application
give him a demo
he�s really into adaptive UIs
the problem is if you adapt to the UI, and the UI is adapting to you
you can easily get out of phase
and if you make things transparent (i.e. signal the adaptations the system is making)
how about doing the Open Mind semantic network based on time entered (so that stuff entered together is clustered together)???
reification vs 2nd order???
lazy evaluation???
what�s CycL???
regression test???