An improved hash function inspired by the fly hashing for near duplicate detections

This paper addresses the problem of improving the fly hashing [1] that is a high-dimensional hash function based on the fruit fly olfactory circuit. The encoding of fly hashing only uses sparsely addition operations instead of the usual costly dense multiplications, and thus results in efficient computations which is important for near duplicate detection tasks in large-scale search system. However, the firing rate based winner-take-all (WTA) circuit of it is neither biologically plausible nor energy saving, and if this circuit is taken into consideration, theoretical results of locality-sensitive are no longer strong. To improve the fly hashing, we proposed a locality-sensitive hash function based on random projection and threshold based spike-threshold-surface (STS) circuit, and both of them are biologically plausible and can be computed very efficiently in hardware. We also presented a strong theoretical analysis of the proposed hash function, and the experimental result supports our proofs. In addition, we performed experiments on datasets SIFT, GloVe and MNIST, and obtained high search precisions as well as fly hashing with less time to consume.