Reading more of the material, and the published paper on solving the Raven's Progressive Matrices, I'm not convinced that the RPM situation is as impressive as it seems. It reminded me of Bart Kosko's writings from around 20 years ago on using a Neural Network to find a ruleset to be used by a fuzzy logic processor to back up a tractor-trailer truck to a loading dock. And the HD/VSA reminded me of the fuzzy logic processor, though of course they are not identical. Here's something I just wrote a friend about it:
"I'm not sure that the Raven's Progressive Matrices test is as big a deal as it sounds, though. It can be solved basically by a large lookup table, which gets constructed by a neural network. The NN is not necessarily an intrinsic part of their system, but it is a way to recognize the problems to be solved. In this case, there are a finite number of permutations of the matrices, so if the NN can recognize the input arrangement, the system can look up how to proceed. This new system can be looked at as a way to encode the permutations and an algebra for computing the lookup. Back in the late 90s or early 2000s, Bart Kosko wrote a couple of interesting books on fuzzy logic and NNs. One of the then-classic tasks was to back a tractor-trailer truck up to a loading dock, in computer space, of course: they weren't hooked up to a real truck. He showed that with the right collection of fuzzy rules and a fuzzy processing algorithm, the truck could be backed up effectively and robustly. And guess what? One way to discover the rules was with a NN. Given the rules, you didn't need the NN to park the truck, just the fuzzy ruleset and processor. It sounds so similar to the present work. In fact, the two main operations are correlation and projection. The rules for doing so are a little different from the fuzzy logic case, and the values in the vectors' cells are binary instead of (possibly) continuous. But that doesn't seem so different to me. One difference with the current work is that you can do symbolic math with their sparse vectors. In simple cases you can sometimes find the answer using an algebra without actually evaluating anything. That's way cool, but I think for most large cases you would have to do numerical evaluation. " -- You received this message because you are subscribed to the Google Groups "leo-editor" group. To unsubscribe from this group and stop receiving emails from it, send an email to leo-editor+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/leo-editor/92f1fcc6-3702-4093-af37-bc468d1fec95n%40googlegroups.com.
