Schedule of Topics
Wiki Stuff |
Topics /
RationalisticAssigned Reading
Additional Resources and Readings
ResponsesWhat Role for Computers' "Intelligence?" - Matt Garr Andy Clark's Mindware provides an interesting juxtaposition to the discussion of paradigms, as it seems there is no mature paradigm guiding the discussion of whether computers can be intelligent, or experience consciousness. It is clear that the question of whether computers can ever experience the richness of consciousness so-called “meat machines” still belongs largely in the realm of philosophy. In the rubric of Thomas Kuhn, there is still no consensus upon the basic facts and approaches to understanding, so, absent a paradigm, it is unclear how to begin collecting data or completing further studies that push us closer to a proper model, let alone a real understanding. Understanding whether silicon-based devices can ever provide the holistic understanding that people seem to have about their world is, however, an extraordinarily important question for computer scientists to attempt to understand. An answer to this question would inform computer scientists as to which problems they can realistically hope to solve, and in what manner. With my own background, as a computer scientist, and a commercial construction superintendent, I have seen the promise (and challenges) of the discipline of artificial intelligence, and approach intelligent construction systems with some skepticism. Much research in construction engineering and management currently focuses on codifying rules of construction and then using these to optimize schedules. However, in the real world construction companies tend to simply ask seasoned superintendents to create construction schedules. Construction firms seem to firmly believe that these individuals have the experience and know how that it takes to get the job done. The corollary is that they feel a computer could never know what it takes to get a major job done. In truth a construction job is not just a set of parts going together. The humans in the organization rely on relationships, shared experiences and gut feelings to move towards a goal. Such a rich set of rules can only be achieved, it seems, through a real-world experience and is why many in the construction industry look disparagingly upon what comes out of the so-called ivory tower of academia. Because it seems doubtful that a computer can (in the foreseeable future, anyway) can rely on a rich real-world experience, it seems that in many cases a computer is better used as a tool that allows a human actor additional tools with which to codify and communicate his or her own knowledge and intelligence, rather than replacing that person entirely. It seems augmented intelligence might be more appropriate than artificial intelligence. One remarkably human phrase kept coming back to me when reading Clark's piece. When a person seems clueless in some social context, a common refrain is “you need to get out more!” Implicit in this adage is a sense that intelligence, whether social or otherwise, is directly related to one's accumulated experiences in the real world, both social and physical. The human experience, it therefore seems, can not be divorced of the human body, with its unique set of sensors and physical capabilities. Until computers can have the richness of the human experience, it seems, they will forever rely upon humans to inform them about the outside world through formal rules. This very fact puts computers at a tremendous disadvantage and in an ever-subordinate role to the humans who must forever teach them. Is the goal really a cognitive computer? - Robert Graebert I very much enjoyed reading the first chapters of Mindware as it was my first exposure to cognitive science. The question I pose is this: Do we really want/need computers to be cognitive? I see a lot of quality in making a computer as intelligent as it needs to solve specific tasks. There are still a lot of problems that need to be resolved to make the power of computers as accessible to the “meat machine” to give expert advice in a field and not just be “dumb” number cruncher. To give one example: The success and failure of Wolfram|Alpha; this is a system that already has a lot of symbolic and quantitative data about certain domains. As long as it understands a question well it will give you answers that could qualify it as an export in its fields. However, it typically does not understand what you are trying to ask unless you use a well-phrased question. With this problem eventually overcome by potentially using the Clark’s discarded methods (artificial neural networks and symbols systems) we will end up with an intelligent answering system for questions for many computable questions. Similarly for other applications the desired end result might not be a cognitive system but rather many individual systems that in themselves represent expert knowledge based on the described methods. So, additionally to asking if cognitive systems can exist outside the well known “meat machine” (and I see no reason why meat should be the only conduit for this; especially in the light of Section 1.3) I think we need to ask ourselves what we will gain from this and if the tools we discard for reaching a robust cognitive system might not produce systems that are already very powerful to us in themselves. Mindware Misses - MFSchar After class comment -- Nate I'm having a little trouble seeing any problems or limitations with the rationalistic tradition, but I'm patiently expecting trouble. When people say "why cant we build computers that act like humans?" I like the answer that we just need to have faster/bigger programs. We haven't really talked about abstractions yet. We can have symbols on a low level and more complex "objects" on a higher level. To say that a chess playing computer is not rational because it will play chess in a burning building is fair depending on your definition (or interpretation/frame if you prefer). But really what if the only difference between that chess playing machine and a "rational" human is what high level objects (abstractions) it is aware of. To make that chess playing machine run away from the fire all we have to do is give it a thermometer and some wheels and tell it to run away when it gets too hot. Is it possible that the only difference between rational human behavior and machines is simply the size of the program? Nate Nate -- on the first part of this comment about how low level the symbols are: As the symbols involved are lower level, it can sometimes be unclear whether they are interpretable or straightforwardly combine in those ways. As a comparison point, connectionism (which comes up more later in the course, see the Dreyfus & Dreyfus article on Topics.PhenomenologicalCritiqueOfAI) works just fine with a computational theory of mind, but it can be unclear what the symbols represent. -- Dean I find it hard to follow, at times, what exactly the difference is between a high level object and a symbol. If symbols are the fundamental unit of cognition, then what are they exactly. Is the "chess board" a symbol, or an object? Maybe each square is a symbol? Does that even matter? The bit about connectionism goes a little ways towards explaining the distinction in terms of what the fundamental units of cognition might be, but even that theory allows for higher-level concepts that are ill-defined. I wonder how a conversation about AI goes without the idea that symbols can be combined to make symbols-of-symbols, or these high level concepts. Or how it goes if you cannot agree on the nature of the symbols in the first place. It seems to me that you must accept these premises to have conversations about the nature of artificial intelligence, but we shall see as the quarter rolls on. -- Chris |