Symbolic Debate in AI versus Connectionist - Competing or Complementary - Essay Example

SYMBOLIC VERSUS CONNECTIONIST: COMPETING OR COMPLEMENTARY Are the Symbolic and Connectionist Approaches to Artificial Intelligence Competing or Complementary Author Institutional Affiliation Are the Symbolic and Connectionist Approaches to Artificial Intelligence Competing or Complementary The symbolic versus connectionist "debate" in AI circles has been on for too long now: it is time we arrived at a consensus, as it were. Consider as an example that archetypal problem of AI, that of developing a chat-bot. If we took a rule-based approach, we end up with a system that is too inflexible, and which produces the kind of responses that the stuff of many jokes and much ridicule; if we were to follow the "learning" approach, we might find that certain items of vocabulary not in the training set would stymie it, leading to no response. Many problems in AI are like this one, where members of each "camp" would like to do it their way. If the aim of human-oriented AI is to develop common sense, an extreme example of the purely symbolic approach is to be seen in the Cyc project (Doug Lenat). Here, common-sense rules inferred from the everyday world are hard-coded into the system such that it will be able to handle any type of situation. And it is in this "extremely symbolic" approach that the worst failures of that approach will probably be seen: forget one fact, and the system crashes, with nothing to lean back on. On the other hand, best-suited to the connectionist approach are models of the brain at the micro-level. The brain is, after all, a neural network-literally. The problem here is that we get a working model, but with very little description of what is actually going on inside, and the question begs to be asked (by connectionists, of course): why model it if it cannot be explained The natural thought is that there must be some way the two systems can "co-operate." Consider an interesting problem, one that may seem far-fetched but which is good enough to serve as an example: that of nonsense translation, as in "English French German Suite," quoted in Gdel, Escher, Bach: An Eternal Golden Braid (Douglas Hofstadter, 1979, page 366). Here, a translation into German by Robert Scott of Lewis Carroll's Jabberwocky is presented. The English stanza 'Twas brillig, and the slithy toves All mimsy were the borogoves, And the mome raths outgrabe." Gets translated into the German as "Es Brillig war. Die schlichten Toven Und aller-mmsige Burggoven Die Mhmen Rth' ausgraben." Consider "outgrabe": how would one "translate" it into German It turns out that "out" is "aus" in German, and "grab" sounds perfectly German; add to that the common German "-en" suffix, and one gets "ausgraben." Similar principles apply to the translation of all the nonsense words here. A connectionist system would be ideally suited to "get the feel" of English and German. To anyone conversant with both languages, it is patent that "mimsy" deserves to be "translated" as "mmsige"! But "feel," as it were, simply does not exist in the connectionist paradigm. But what about the grammar "'Twas brillig" cannot be "Es war Brillig"; it must be "Es Brillig war." (Word order.) Who is to provide these rules It should be amply clear that only a co-operation of the two paradigms can plausibly take on this task of nonsense translation, which, though not a pressing problem in the real world, is indicative. Think, now, of a system that checks English sentences for grammatical correctness. The most obvious way to do it would be via a rule-based system, because a grammar is, after all, rules. However, this does not work in the real world. Think about the cardinal rule of English grammar, that every sentence must have a verb. How grammatically correct, then, is the sentence "Why not" And what about this perfectly grammatical sentence Given a standard book, a recursive neural network would be able to measure with reasonable accuracy whether a sentence is grammatically correct or not. But what if the test text is in a different language of discourse The neural net would undoubtedly flounder-it needs rules. Only a combinational approach would work, it seems. We have then two examples of problems that seem ideally suited to one of the two approaches, but which would actually stand to gain heavily by an infusion of the other. The same sentiment is expressed in Dimensions of Neural-symbolic Integration - A Structured Survey (Bader and Hitzler, 2005): "During a training process neural networks acquire expert knowledge about the problem domain, and the ability to generalize this knowledge to similar but previously un-encountered situations The knowledge obtained however, is hidden and not directly accessible for analysis, reuse, or improvement. Likewise, in situations where partial knowledge about an application domain is available before the training, it would be desirable to have the means to guide connectionist learning algorithms using this knowledge." Coming to our own selves, we find that we ourselves are hybrid systems. At least, that's what we must infer from the quotes of the experts: "We find it particularly hard to use our language skills to talk about the parts of the mind that learned such skills as balancing, seeing, and remembering, before we started to learn to speak." (Minsky, The Society Of Mind, Section 6.12) Indeed, we might extrapolate and say that our symbolic selves cannot even learn to do such things! And language skills on the one hand (symbolic) and balancing and remembering on the other hand (connectionist) are both essential components of our intelligence. If our brains are thought of as connectionist simply because they are neural networks, that thought should be put to rest by the following from Logical Versus Analogical or Symbolic Versus Connectionist or Neat Versus Scruffy (Minsky, 1991): "I suspect that our societies and hierarchies of subsystems have evolved ways to evade the problem (of network opacity) by arranging for some of our systems to learn to model what some of our other systems do." A little self-analysis should suffice to tell us that "we" are connectionist-symbolic hybrids. A last word-about consciousness and self-awareness-is in order here. Our consciousness is largely a result of our ability to loop back upon ourselves, as it were. To loop back, there must be a symbolic agent doing the looping. At the same time, there must necessarily be connections between our component selves to enable the looping. If what we're saying is taken as plausible, our conscious structure itself is neither purely symbolic nor purely connectionist. In fact, each by themselves has no meaning in the context of consciousness. We conclude that in artificial intelligence, and in the cognitive sciences in particular, where such things as the origins of our own intelligence and consciousness are considered, there is no place for bickering over the symbolic/connectionist debate-it must necessarily be a fusion. Read More