Greg Detre
Wednesday, December 04, 2002
formal contexts as a Minsky (partially filled in) frame???
multiple frames at a time ~= reasoning in different contexts
solve the frame problem???
MM allows for some messiness � context doesn�t help you there, does it???
what does a NN give you??? sub-features (Smolensky) and it self-organises
are these two important???
causal model (i.e. discussion with Hugo) as necessary in context too???
partial, approximate and perspectival
MCS � multi-context system
reasoning as an aspect of context
principles of locality and compatibility
PLC doesn�t formalise localised vocabularies
Ted: difference between them
PLC it�s the switching between contexts that is the reasoning
MCS it�s the interface between the contexts
why in MCS isn�t there a big unifying world (-context)???
Ted wants to keep all the implicit contextual assumptions implicit
but the price of not having to enumerate all the assumptions is brittleness, i.e. it can only respond in the way it�s been programmed (and those responses contain the assumptions implicitly), right???
Andrea: what you don�t explicitly say you can�t switch out of
this is just choosing a different point along the trade-off curve between the amount of work you have to do enumerating and brittleness, right???
his idea can be reduced to a formal system (even the order can be defined in terms of antecedents)
all he�s saying is that you only provide useful rules, rather than lower-level rules from which you can generate them
an FSM can�t change � but if you�ve inferred more or less everything useful/likely from your set of propositions, then it�s no difference
if you make it a probabilistic FSM you could incorporate contextual information from the variables
Napier: PLC requires you to know what context you�re in
does LCS/MCS help here???
has the cyc
12d context mechanism been added??? does this mean that the database is more or
less entirely self-consistent now???
������� does more or less every assertion now
reside in a context???
������� do all questions have to be reified???
if there was a customised commercial
application based on cyc, would that be a 'topic'???
let�s say:
you�ve got 2 years of the Boulder house�s light switch and heater data
you provide a fixed ontology
maybe if you tell the house what context you�re in once an hour, and it has high-dimensional sensor data, then it could start to map your current activities to sensor data
how would you generate new contexts??? when current sensor state doesn�t fit with a stored attractor closely enough
if you come in at a certain time of the day and turn on one light switch, it would guess what the complete state of the house you want is
you could do this easily for a UI
it�s easy to get a log from the Mac � recognise patterns of context
her friend doing this wants your computer to know about quality of life issues, e.g. workload, stress
you�d combine the operational context and the user model into a context, e.g. learning to use MS word, learning a new programming toolkit or function, messing around on slashdot, doing a complicated cad drawing writing a paper,
what do you do with this information???
prioritise ram
make some icons more prominent
load up reconnaissance agents
build up user model of competence
help the user use the computer more effectively
how much information to present to the user
different types of user
suggesting more effective ways of working
cross-application help systems
whether or not to make recommendations
setting icq status
Sims doesn�t provide enough of a log