A General and Efficient Querying Method for Learning to Hash

Pm-LSH

“…Probing Sequence Based (PS) Approaches. The representative PS methods include Multi-Probe [22,23] and GQR [20] that use a carefully derived probing sequence to examine multiple hash buckets that are likely to contain the nearest neighbors of a query point. Unlike the basic LSH that builds L hash tables and checks only one hash bucket in each hash table, PS probes multiple nearby buckets in order to achieve higher recall with fewer hash tables.…”

“…, ρm] is computed as follows. To evaluate the performance of our estimator in Lemma 2, i.e., L2 = r (the same as our estimator when m is fixed), we compare it with other distance estimators: L1, QD [20], and Rand (assign a random value). We randomly sample a small dataset that contains 10K points from the Trevi dataset [21] and choose 100 points as query points.…”

“…Datasets and Query Sets. We use seven real datasets: Audio, Deep, NUS, MNIST, GIST, Cifar, Trevi, which are used widely in existing work [11,17,20,21,31] on LSH. Table 3 reports the dimensionality and key statistics of the datasets: Homogeneity of Viewpoints (HV [7]), Relative Contrast (RC [16]), and Local Intrinsic Dimensionality (LID [2]).…”

“…LSH uses predefined hash functions without considering the underlying dataset, while L2H learns tailored data dependent hash functions. Many learning algorithms have been proposed, such as iterative quantization (ITQ) [13], and generate-to-probe QD ranking (GQR) [20].…”

“…A query is answered by examining the points that are hashed to the same bucket as the query point, or to similar buckets. Based on their main ideas, we classify the mainstream LSH methods into three categories: 1) Probing Sequence based (PS) approaches [20,22,23]; 2) Radius Enlarging based (RE) approaches [11,17,33]; and 3) Metric Indexing based (MI) approaches [31]. PS approaches use a carefully derived probing sequence to examine multiple hash buckets that are likely to contain the nearest neighbor of a query.…”