Schedule of Topics
Wiki Stuff |
Topics /
PhenomenologicalCritiqueOfAIAssigned Reading
Additional Resources and Readings
Comments - Before ClassForest - Closed loop: from Descartes to Winograd and back againThe background of artificial intelligence provided by Haugeland shed light on the "family tree" of past contributors and their points of departures. In my literature I found this tactic useful to help provide insight into the differing branches of model-based research. The result was a tree with branches that expanded from a professor that propositioned a particularly ground breaking theory, through their thesis advises and to their thesis advises, to form the domain of a paradigm. At several points the tree formed a circle as cross-pollination occurred from differing branches. The point is it provides insight on the background. As for Haugeland, the path of the first empirically model-based science with Copernicus moved to mathematical modeling and the nth dimension. The principle origin of AI was illustrated to have been Hobbes, a particularly gifted political scientist but poor mathematician. from there Haugland lays the path that leads through the development of artificial intelligence principles that essentially still hold today, such as, generality and rules-based experimentation. At the same time as Descartes establishes these principles, the proposition is advanced that all this will result in naught. This final proposition is the crux of Winograd's conclusions drawn in Chapter 8 of "Understanding and Being, Computers and Representation, Computation and Intelligence." Winograd then supports this assertion in Chapter 9 with examples of well resourced projects that repeatedly have failed. The best example being the "How long have you been swallowing poison" response from DOCTOR! It is not explicit, but Chapter 10 appears to be presenting a forecast of the next round of failures. A check on Wikipedia confirms that fifth generation computing , was in fact, a failure. The realization that Professor Winograd is laying out a well constructed argument for the scientific disproving of artificial intelligence required digging deeper into more recent developments. A review of work published by Rodney Brooks (2001) provided the update, beyond the "indictment of classical AI". "At the heart of this disappointment lies the fact that neither AI nor Alife has produced artefacts that could be confused with a living organism for more than an instant." Rodney Brooks lays the facts bare, advances in computing hardware has increased computation of chess moves from 2K per second to 2,000K per second, and increased the frame rate of object recognition from 1 every ten minutes to 30 per second, allowably resulting in great increases in perceived capabilities, but again nothing resembling life. For my own research this realization, is both appreciated and despised. First, I can remove my ridiculous obsession with the dream of fully autonomous project planning and control. second, I rather liked that obsession. The path forward is provided by both Winograd and Brooks. The advances made in the last 500 years provided by Brooks, while validating Moore's law on capabilities, seems to be flat within the realm of theory as Descartes seems to be still the leading theorist. Winograd provides the avenue of advance as is apparent by the movements made by other virtual reality researchers, is to facilitate the use of machines in their capacities most appropriate. Nate"The most significant advances in computer science in the coming decade will be those facilitating [human-computer] interaction"-- I agree completely. Machines with human-level intelligence seem a long way off. As pointed out its not impossible that some program could grow and evolve to have human-level intelligence. I like the analogy to the search for the philosophers stone (by the way, I had no idea we could actually turn lead into gold). If we finally do get machines with human-like brains I'm not so sure the successful approach will be so distinct from the "evolution of structure" approach, well... at least not as distinct as alchemy and particle acceleration, but who knows. Theres been a lot of emphasis on the context and background of words when analyzing their meaning, so much so that it does seem obvious that any system that were to account for the meaning of words as humans know them would not be programmable by one or even a small group of people unless it had an evolving structure. Although, I will concede that the arguments for the limits of the speed of artificial evolution are strong. It does indeed seem that the best thing we can do is build machines that are more appropriate for their given tasks, but maybe by solving these HCI problems we will naturally move closer and closer to solving some of these human-level intelligence problems. Also, its still fun to think that when doing your programming homework your possibly working with the same stuff that consciousness is made of : ) Searle's argument is fun, but ultimately I agree, its really a language puzzle in disguise, or perhaps a veiled argument for dualism. Ultimately its practical in the sense that its thought provoking, but does it lead us to build anything new and useful? Also, some questions I had when reading: are these "Evolution of Structure" programs formal systems? Are there alternative definitions of a computer to Turing's Machine? Perhaps one where the rules are not static--one with an evolving set of rules? Artifacts and Tracing Meaning - Jeff WearWhat does Haugeland have against chess? Twice in his discourse on semantics, he asserts that chess games are meaningless (96-97). I think that Bobby Fischer, and many anthropologists, would tend to disagree. Haugeland clearly believes that it is appropriate to interpret linguistic text - to "make sense of [that body of symbols] as symbols" (97). In linguistic texts, we may discover coherence (order and sense) - and perhaps other sorts of texts; Haugeland says that "if [coherence] is there to be found, then there's no further question as to whether the tokens are 'really' interpretable.' Well then, let us seek coherence in chess. Clearly chess is ordered: in the course of play, each state (configuration of pieces on the board) transitions to the next state according to precise rules. Haugeland does not deny chess its logic, but nevertheless dismisses its interpretation - "chess games ... don't 'say' anything - even though they're not random." (97) Can we make sense of chess? What meaning could chess convey? The answer is simple: chess conveys the intentions of the players, and as such, is meaningful. Chess games are more than structured, they are "constructed" by us, a distinction Haugeland himself makes. A game, considered at any one moment, is a history of decisions made by the players individually and in interaction with each other, and this history is purposeful. Not every arrangement of pieces, even states that have been reached via valid applications of the rules of chess, could be construed as "making sense" within the context of chess' aims. Imagine a board where both players have failed to take each other's queen despite multiple opportunities to do so: this board would be somewhat implausible. Haugeland may yet protest, as he does on pg. 96, that chess cannot be "construed as 'saying' anything [e.g. about the players' intentions, as I describe above] by virtue of [chess'] structure and the 'meanings' of [its] components." It does seem that chess, as an ordered text, is unlike language in that individual pieces do not admit semantic interpretation. But then again, do individual words "stand alone"? The word "the" signifies nothing in and of itself. In Haugeland's examples of cryptograms, he admits that ordered texts may not support interpretation absent a certain amount of internal structure. In the same way as words acquire meaning at the level of phrases and sentences and in relation to other words, chess pieces acquire meaning at the level of formations and games and in relation to other pieces. Chess cannot be interpreted as 'saying' anything in the absence of a mind familiar with its operation or keenly observant of its workings, but then this is also true of language. Both are communicative systems, or can be regarded as so, because their architects have invested them with meaning in a fashion which is not only ordered but purposeful. It is this intentionality that makes it possible to reasonably construe these systems as meaningful. Chess, circuit boards, etc. are not mere things, they are artifacts, to be distinguished from naturally ordered constructs like molecules, because humans have given them a structure which preserves their meaning beyond the act of construction. Luke - Lost in translationI must admit that this I felt a bit lost amongst the many details of this week’s readings. Perhaps as I gain familiarity with the discussion I will be able to better distinguish the different stances and their arguments, but at the moment I don’t think I’m seeing the greater arc. I will however mention a few things that I noticed or appreciated. It seemed to me that the majority of the discussion has to do with defining terms: What does it mean to understand? What would it mean for something/someone to be intelligent? What is the meaning of “to mean”? I particularly liked the definition of intelligence proposed on p98 of Winograd and Flores: “The essence of intelligence is to act appropriately when there is no simple pre-definition of the problem or the space of states in which to search for a solution.” This description encapsulates what, I think, is meant or intuited in the common usage of the term. Speaking of intuition it seemed that Searle in his Chinese Room discussion made many appeals to our intuition of what it means to be something “like us” that understands, but he did not make explicit what characteristics we have that are required. He used the term “causal powers” a number of terms to denote what brains can do that computers might not, but in this case we lack even an intuition of what this means. Once again I am impressed by the illuminating quality of Heidegger’s philosophy. His description of the “blindness” inherent in our use of a set of properties to characterize an object (which is present-at-hand) is a great way to point out the limitations of representation. After Class Comments - ForestThe illustration by Professor Winigrad of the five traditions of AI helped very much to frame the topic. That there is more than just the "neats" and "cleans" and how this course material fits within the tree helped me better understand the course. That Phenomenology is a form of study, equivalent to other AI research, had somehow been lost on me until this point. I hope to not sound too dense, but better to "get now" now than never. The traditions of AI and how this fits with cognitive science is still a mystery to me. The next question I have is how many sub-traditions are there to these five traditions. And do these five traditions encompass the entire domain of AI research. Last, is these a concerted effort to disprove AI? After Class Comments - Mark ScharI agree with Forest, the highlight was Professor Winograd's "off the cuff taxonomy" of AI. My notes suggest four branches: Computational (GOFAI), Neural (fMRI), Pragmatic (natural language translation) and Phenomenological (Dreyfus, Searle, et al). This was prompted by the discussion of "neats vs. scruffies" which helped place the conversation in context. Also, I enjoyed discussiong the readings with Forest ahead of time, as it helped put everything in perspective. In recent weeks, I've come across resources who are working on versions of AI which may be interesting to the class. I attended a lecture by Kwabena Boahen of Stanford's Bioengineering Department. Kwabena talked about the work in his lab to model neurons and neuron networks using transistors - something he calls the Neurogrid. At this stage the can model process up to one million neurons and six billion synapses, which approximates the neural capacity of a honeybee. Very interesting. A second lecture was by Professor Randy O'Reilly of the University of Colorado - Boulder. O'Reilly is working on a model of human vision called "EMER" which is modeling the V1, V2, V4 visual systems and the associate learning function. There is a very cool process for binocular foviation used in 3D object identification. They are modeling dynamic learning by capturing the action potential relationship between neurons under "correct" and "incorrect" conditions. O'Reilly believes he is modeling Lakoff's "emergent cognition" hence the name for his model - "EMER." If interested, it's worth joining the Mind Body Cognition lab as an affiliate and get the notices of these speakers. Finally, there is a great article in this month's Scientific American Mind on What does a smart brain look like? It says that new research shows there may be several/many pathways for cognition - where differences outweigh similarities. This will throw a monkey wrench in current neural modeling which looks for consistencies, not differences. Good stuff! |