Greg Detre
Tuesday, 29 January, 2002
3D error landscape
propositional modularity, ICS, advantage of connectionism = self-org???
competence/performance (see Cohen in thesis)
sub-doxastic
linguistics vs connectionism
connectionism as glorified statistics???
the top level of mind can be seen as a VM (conscious rule-interpreter) of cultural knowledge (that resides outside a single individual), often formluated linguistically
VM as intuitive pre-processor (i.e. carries out practically all skilled performance)
sub-symbolic ≠ neural
stochastic c.= randomness
soft/hard vs discrete/continuous
computational temperature c.= simulated annealing
harmony theory
thermodynamics
can classical models self-organise at all???
crucial to the success of self-organising connectionist systems is the idea of forming a SPACE
and I think divorcing the algorithm from the concepts is necessary to do this � a non-connectionist approach which encoded numerically and computed sub-symbolically could be potentially as powerful
and, given that connectionist self-organising requires so much fine-tuning, and humanity uses DNA, broadly pre-specifying/constraining the symbolic module isn't cheating to a much greater extent
both can exist within the Godel-numbering/function gen/NN system I�m planning too
our minds (and even, I think, pretty much �mind� in general) cannot be formally specified at the cognitive level
why do I think that this is a property of all minds??? what about minds that live ina� much cleaner environment, or have more low-level processing or something??? somehow I feel that there is an optimum window/balance of chaos + order (like in A-life and CAs) necessary for both powerful abstraction + analogy
how might conflicts tewen philosophy of mind and connectionism drive each other forward in the future???
in the future, tolerance for biologically implausible dynaimcal systems will �, though perhaps not fully until NNs approach brain-standard
the field of AI will have to be very sensitivce to when we do and when we don't want to closely model human-like mental processes
can we measure synaptic weights empirically yet???
surely yes, e.g. LTP etc.
how might we know the inside of the brain???
with nano??? imaging???
can is Smolensky saying that the system can be wholly described at the sub-symbolic (rather than neural) level???
yes
intuitive vs rule-processing???
surely even the sub-conceptual is technically semantic if you�re prepared to use non-linguistic/boring truth conditions for each neuron???
if so, then how can we know what we mean by semantic at all???
on Smolensky�s account, are the rule-interpreter + intuitive systems functionally separate???
conscious as large-scale, stable(/stereotyped???) patterns of activity???
even though the rule-interpreter is fully-specifiable at the conceptual level, the intuitive linguistic device is not
symbol � corresponds/represents something in the external world
manipulated syntactically (i.e. by shape???)