David Barbour <[email protected]> writes: > Well, communicating with genuine aliens would probably best be solved by > multi-modal machine-learning techniques. The ML community already has > techniques for two machines to "teach" one another their vocabularies, and > thus build a strong correspondence. Of course, if we have space alien > visitors, they'll probably have a solution to the problem and already know > our language from media. > > Natural language has a certain robustness to it, due to its probabilistic, > contextual, and interactive natures (offering much opportunity for > refinement and retroactive correction). If we want to support > machine-learning between software elements, one of the best things we could > do is to emulate this robustness > end-to-end<http://awelonblue.wordpress.com/2012/05/20/abandoning-commitment-in-hci/>. > Such things have been done before, but I'm a bit stuck on how to do so > without big latency, efficiency, and security sacrifices. (There are two > issues: the combinatorial explosion of possible models, and the modular > hiding of dependencies that are inherently related through shared > observation or influence.)
As well as purely statictical ML, I think John Pollock's defeasible reasoning is an interesting model. Results are given alongside a justification/argument for why. These justifications can be defeated later, causing the result to be revised. This is reminiscent of proof-carrying-code, but using non-monotonic logic. It is also similar to "Cromwell's Rule" 'think it possible that you may be mistaken', which current software either ignores, or handles simplistically with exceptions. Worlds offer a better approach to this, since the system isn't left in an inconsistent state. Delimited continuations also look interesting for this kind of undo-able computing. There may be interesting parallels to reversible computation as well: if we allow computations to be undone/revised then they can't destroy any information. Another thing that comes to mind from this conversation is CosmicOS ( http://people.csail.mit.edu/paulfitz/cosmicos.shtml ). Regards, Chris _______________________________________________ fonc mailing list [email protected] http://vpri.org/mailman/listinfo/fonc
