The Research For Anchor Hash In Large-scale Machine Learning

When searching in large-scale databases efficient, hash has become increasinglypopular. However, for shorter codes to learn, to those of better performance searchalgorithm, is still a challenge. Moreover, the data in the real world is almostlow-dimensional, which at the time of the capture of the neighboring areas shouldconsider getting.Semi-supervised learning(SSL) has an important impact on theuniversal application of machine learning and image recognition.In these practicalapplications,a frequently encountered situation is that only a small part of the tag dataavailable and most of the the data is not marked.Tagged data acquisition time istypically difficult and costly,but not able to automatically collect data marked and costvery little.In a large number of SSL method,based on the figure of semi-supervisedlearning(GSSL) is attractive because it can be easily implemented and generallyshows a closed solution.With the development of the network,and now we can collectlarge amounts of unlabeled data,then the demand for large-scale SSL.Despite theadvantages GSSL draw in practice,due to the great cost calculation step,from a largeclassification diagram illustrating the structure and training,most GSSL methodperformance in terms of the data size is very full.In this paper, we propose a method that based on hashing, it can automaticallydiscover the inherent data near the body, which can learn the proper compact code. Tocomplete this approach, we use the "anchor map" to get the low-level adjacencymatrix relatively easy to handle. This method provides a fixed time, by hashingcalculate new data points Laplacian image feature vector characteristic functionobtained. Finally, we describe a classification threshold learning process, where eachfeature has multi-function bits, thus getting a higher search accuracy. Experimentscompared with other advanced methods on two data sets presented.