--- Derek Zahn <[EMAIL PROTECTED]> wrote: > At any rate, if there were some clearly-specified tests that are not > AGI-complete and yet not easily attackable with straightforward software > engineering or Narrow AI techniques, that would be a huge boost in my > opinion to this field. I can't think of any though, and they might not > exist. If it is in fact impossible to find such tasks, what does that say > about AGI as an endeavor?
Text compression is one such test, as I argue in http://cs.fit.edu/~mmahoney/compression/rationale.html The test is only for language modeling. Theoretically it could be extended to vision or audio processing. For example, to maximally compress video the compressor must understand the physics of the scene (e.g. objects fall down), which can be arbitrarily complex (e.g. a video of people engaging in conversation about Newton's law of gravity). Likewise, maximally compressing music is equivalent to generating or recognizing music that people like. The problem is that the information content of video and audio is dominated by incompressible noise that is nontrivial to remove -- noise being any part of the signal that people fail to perceive. Deciding which parts of the signal are noise is itself AI-hard, so it requires a lossy compression test with human judges making subjective decisions about quality. This is not a big problem for text because the noise level (different ways of expressing the same meaning) is small, or at least does not overwhelm the signal. Long term memory has an information rate of a few bits per second, so any signal you compress should not be many orders of magnitude higher. A problem with text compression is the lack of adequate hardware. There is a 3 way tradeoff between compression ratio, memory, and speed. The top compressor in http://cs.fit.edu/~mmahoney/compression/text.html uses 4.6 GB of memory. Many of the best algorithms could be drastically improved if only they ran on a supercomputer with 100 GB or more. The result is that most compression gains come from speed and memory optimization rather than using more intelligent models. The best compressors use crude models of semantics and grammar. They preprocess the text by token substitution from a dictionary that groups words by topic and grammatical role, then predict the token stream using mixtures of fixed-offset context models. It is roughly equivalent to the ungrounded language model of a 2 or 3 year old child at best. An alternative would be to reduce the size of the test set to reduce computational requirements, as the Hutter prize did. http://prize.hutter1.net/ I did not because I believe the proper way to test an adult level language model is to train it on the same amount of language that an average adult is exposed to, about 1 GB. I would be surprised if a 100 MB test progressed past the level of a 3 year old child. I believe the data set is too small to train a model to learn arithmetic, logic, or high level reasoning. Including these capabilities would not improve compression. Tests on small data sets could be used to gauge early progress. But ultimately, I think you are going to need hardware that supports AGI to test it. -- Matt Mahoney, [EMAIL PROTECTED] ------------------------------------------- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604&id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
