Unsupervised deep neuron-per-neuron hashing

摘要

Hashing has been widely applied to approximate nearest neighbor search for large-scale multimedia retrieval. A variety of hashing methods have been developed for learning an efficient binary data representation, mainly by relaxing some imposed constraints during hash function learning. Although they have achieved good accuracy-speed trade-off, the resulting binary codes may fail sometimes in adequately approximating the input data, thus significantly decreasing the search accuracy. In this paper, we present a new Unsupervised Deep Learning Hashing approach, called Deep Neuron-per-Neuron Hashing, for high dimensional data indexing. Unlike most existing hashing approaches, our method does not seek to binarize the neural network output, but rather relies directly on the continuous output to create an efficient index structure with hash tables. Given the neural network deepest layer, each table indexes separately a neuron output, capturing in this way a particular high level individual structure (feature) of the input. An efficient search is then performed by computing a cumulative collision score of a given query over all the neuron-based hash tables. Experimental comparisons to the state-of-the-art demonstrate the competitiveness of the proposed method for large datasets.