Compact Hyperplane Hashing with Bilinear Functions Wei Liu (Columbia Columbia), Jun Wang (IBM IBM), Yadong Mu (Columbia Columbia), Sanjiv Kumar (Google Google), and Shih‐Fu Chang (Columbia Columbia) June, 2012 Point‐to‐Point Search vs. Point‐to‐Hyperplane Search normal vector t nearest nearest neighbor nearest nearest neighbor hyperplane h perplane query point query Exhaustive point‐to‐hyperplane p yp p search costs linear time. Hashing hyperplane queries to near neighbors, achieving sublinear 2 time or faster. Hashing Principle: Point‐to‐Hyperplane Angle 1 1 ‐1 ‐1 ‐1 1 The ideal neighbors ┴ w The ideal neighbors ┴ 3 Proposed Idea: Bilinear Hashing Proposed Idea: Bilinear Hashing • A A bilinear bilinear hash function outputs the same bit for parallel hash function outputs the same bit for parallel inputs and tends to output different bits for ┴ inputs: Bilinear‐Hyperplane Hash (BH‐Hash) Any input vector: query normal w or query normal w or database point x. random projection vectors 4 A Single Bit A Single Bit x1 u v 1 x2 ‐1 ‐1 Probability? 1 // bin // bin ┴ bin 5 Theoretic Analysis Theoretic Analysis is the highest collision probability till now. Jain 2010 Ours is highest 6 Multiple Bits Multiple Bits u1 u2 v1 v2 ┴ bin1 ┴ bin2 neighbor region = bin1 ∩ bin2 ∩ bin3… 7 Learning Bilinear Projections Learning Bilinear Projections 1 u w 1 -1 x1 u w 1 x1 v v Make h(x) yield the same bit for nearly // inputs. 8 Learning Bilinear Projections Learning Bilinear Projections 1 1 w w u x2 1 v u x2 -1 v Make h(x) yield different bits for nearly ┴ inputs. 9 Active Learning Experiments Active Learning Experiments • 20 Newsgroup: about 20K documents from 20 classes. 20 Newsgroup: about 20K documents from 20 classes. • Use Use linear SVMs as base classifiers and run 300 active linear SVMs as base classifiers and run 300 active learning iterations per class. • Hyperplane hashing in each AL iteration: use 16 bits to do hash lookup within Hamming radius 3, and acquire a short list containing neighbors from the found hash buckets; scan the list to return the final neighbor that has the smallest point‐to‐hyperplane i h l di distance. 10 20 Newsgroup 20 Newsgroup LBH‐Hash outperforms exhaustive search. LBH‐Hash returns much better neighbors than the others. 11 Please come to our poster for more details! Thanks! Questions? Thanks! Questions? 12