On 10/10/07, Vladimir Nesov <[EMAIL PROTECTED]> wrote: > Pei, > > (Sorry for a long list of questions; maybe I'm trying to see NARS as > what it isn't, through lens of my own approach.) > > Do you have a high-level description of how statements evolve during > learning of complex descriptions, including creation of new > subsymbolic terms (compound terms)?
Nothing too detail, but you can start from the paper I co-authored with Hofstadter, especially on the discussion on "Compositionality" and "Categorical dynamics". I don't call the compound terms "subsymbolic". > Basic rule for evidence-based > estimation of implication in NARS seems to be roughly along the lines > of term construction in my framework (I think there's much freedom in > its choice, do you have other variants of it/justification for current > choice relative to other possibilities which is not concerned with > applicability to derivation of rules for abduction/induction/etc.?), There is some justification behind the design of every inference rule (and its truth value function), not only abduction/induction. You can find most in the book, and many are also in my other publications. > but I'm not sure about how you handle variations of structures (that > is, how does system represents two structures which are similar in > some sense and how it extracts the common part from them). It's > difficult to see from basic rules if it's not addressed directly. The basic rules (deduction/abduction/induction/revision) ignore the internal structure of compound terms. There are special inference rules that handles the composition/decomposition of various compound structure. Again, they are mostly given by the book. > For > example, how will it see similarities and differences between > 111222333 and 111122223333? Would it enable simple slippage between > them? How will it learn these representations? Yes, the two can be recognized as similar, so the analogy rule can use one as the other in certain situations. > Do you address temporal activation of terms, where term being active > is a temporal statement expressed as relative to current moment, and > learning of structure results from prolonged cooccurrence of its > components? Yes, to a degree, though not in the same way as neural network. I'm sorry that I don't have the time to give a detailed explanation on this topic. > Basic rule seems to require presence of terms at the same > time, which for example can't be made neurologically plausible, unless > semantics of terms is time-dependent (because neuron only knows that > other neurons from which it received input fired some time in the > past, and feature/term it represents if it chooses to fire is a > statement about features represented by those other fired neurons in > the past). It depends on what you mean by "presence of terms at the same time". In NARS, all inference happens within a concept (because every inference rule requires two premises sharing a term), so as far as two beliefs are recalled at the same time, the basic rules can be applied. Whether NARS rules are "neurologically plausible" is not a major consideration for me. NARS is not a brain model. > Why do you need so many rules? I didn't expect so many rules myself at the beginning. I add new rules only when the existing ones are not enough for a situation. It will be great if someone can find a simpler design. > Ultimately all you need are rules for > term formation (for which intersection as starting point seems to be > enough) and term activation given currently active terms (fluid > inference). Is there a basic set which is theoretically sufficient, > although probably requires too much indirect support structures (I > assume that input/output experience is presented as flat conjunction > of active terms)? Maybe, but until I see a concrete design, I cannot be sure. > Why do you need to separately regard operations on > terms and statements (and why statements have any significance in > themselves, other than specific interpretation of underlying term > activation rule)? Not fully separate. Statements in many cases are treated just like other terms. However, since statements are "terms with truth value", they do need special treatment here or there, which don't make sense for other (non-statement) terms. Pei > On 10/10/07, Pei Wang <[EMAIL PROTECTED]> wrote: > > In NARS, the Deduction/Induction/Abduction trio has (at least) three > > different-though-isomorphic forms, one on inheritance, one on > > implication, and one mixed. > > > > For people who don't have access to the book, see > > http://nars.wang.googlepages.com/wang.abduction.pdf , though the > > symbols used in that paper is slightly different from the current > > form. > > > > Pei > > > -- > Vladimir Nesov mailto:[EMAIL PROTECTED] > > ----- > This list is sponsored by AGIRI: http://www.agiri.org/email > To unsubscribe or change your options, please go to: > http://v2.listbox.com/member/?&; > ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=52156872-b152b2
