4 ms·
This seems like a rehash of hashing. sorry for the pun. Neural Networks are, in essence, just really good at compression.
by posterboy 9y ago
This seems like a rehash of hashing. sorry for the pun. Neural Networks are, in essence, just really good at compression.
- noelwelsh 9y agoIf you want to view it that way, it is data dependent hashing. That is the innovation: the mapping from data to hash is learned from the distribution of the data.
- js8 9y agoIt's not really that new - gperf has been doing that for decades.
- eternalban 9y agoPerfect Hashing was crica 1984.
- noelwelsh 9y agoPerfect hashing provably has size that is O(n) in the data being indexed. You can think of this as approximate perfect hashing. You trade off a bit of accuracy for a lot of compression.
- eternalban 9y agoThanks. That [is] what I gathered as well. My OP was merely to note relevant existing results.
- quotemstr 9y agoAll model building is compression. That's the point of building models!
- fooker 9y agoThe innovation is not in the technique itself. Showing the technique is practical in current hardware for current problems is what the paper is about.