Rachkovskij D.A.,1 Misuno I.S.,2 Slipchenko S.V.3
1International Scientific-Educational Center of Information Technologies and Systems, National Academy of Sciences and Ministry of Education and Science, Youth and Sports of Ukraine, Kyiv, Ukraine,
e-mail: dar@infrm.kiev.ua.
2International Scientific-Educational Center of Information Technologies and Systems, National Academy of Sciences and Ministry of Education and Science, Youth and Sports of Ukraine, Kyiv, Ukraine,
e-mail: i.misuno@gmail.com.
3International Scientific-Educational Center of Information Technologies and Systems, National Academy of Sciences and Ministry of Education and Science, Youth and Sports of Ukraine, Kyiv, Ukraine,
e-mail: serge.slipchenko.irtcits@gmail.com.
Abstract. We investigate the properties of randomized binary vector representations with adjustable sparseness formed from the input vectors by projecting them using a random matrix with ternary elements {–1, 0, +1}. We analyze the accuracy of estimating the measures of similarity-difference of the source vectors having a floating-point format by the output binary vectors. Those vector representations can be used for an efficient processing of large volumes of input multidimensional vectors in applications related to search, classification, associative memory, etc. Figs: 6. Refs: 25 titles.
Keywords: random projection, vector representation, sparse binary representation, distributed representation, efficient similarity estimation.