ideas-ai-compressionIdeas

Difference between revision 2 and current revision

No diff available.

http://en.wikipedia.org/wiki/Hutter_Prize

Have a very short token that means 'your first guess is correct', e.g. "i'm not going to explicitly say what goes here, but compute the most probable guess, and that's it". And then an almost as short token that means "your second guess is correct", etc, up to some small number such as 7 (in theory the choice of the ceiling number can itself be optimized). E.g. if you see "at the end _ the", you can infer that _ should be "of".

(see http://rsif.royalsocietypublishing.org/content/early/2012/07/23/rsif.2012.0491.full for example phrases)

Construct a hierarchical ontology. Make assumptions about the statistics of the branching nature of the ontology (power law?).

---

also, it seems to me like one of the best extant compression algorithms, http://en.wikipedia.org/wiki/Lempel%E2%80%93Ziv%E2%80%93Markov_chain_algorithm (the one used by xz, LZMA2), doesn't even do any wavelet transform stuff? That would explain its poor performance on my image raster .mat files.

shouldn't you at least do trend subtraction? i understand that you don't know how numbers are being encoded but surely you can do something that works at least somewhat using current common encoding formats?

often even non-image data will have repeated packets/records of a certain length, a wavelet approach could help there

---

combining those two, we have a spatial template of certainty/uncertainty probability

generalizing beyond space, instead of a template parameterized by space, we get one parameterized by anything

applying my many-linked-state-machine-estimatorsidea, we have many bagged estmators

and also apply my idea to link them with spreading activation

and then also we also have an abstraction operation that allows some of those to orchestrate sequences of others

---

how would you compress an image of snow on a tv screen, where the viewer doesnt care about the exact piciture, but only the general feeling/texture?

---