HashNet: Deep Learning to Hash by Continuation

By Zhangjie Cao, Mingsheng Long, Jianmin Wang, Philip S. Yu. IEEE International Conference on Computer Vision, 2017

724 citations recorded by the source corpus.

Learning to hash has been widely applied to approximate nearest neighbor search for large-scale multimedia retrieval, due to its computation efficiency and retrieval quality. Deep learning to hash, which improves retrieval quality by end-to-end representation learning and hash encoding, has received increasing attention recently. Subject to the ill-posed gradient difficulty in the optimization with sign activations, existing deep learning to hash methods need to first learn continuous representations and then generate binary hash codes in a separated binarization step, which suffer from substantial loss of retrieval quality. This work presents HashNet, a novel deep architecture for deep learning to hash by continuation method with convergence guarantees, which learns exactly binary hash codes from imbalanced similarity data. The key idea is to attack the ill-posed gradient problem in optimizing deep networks with non-smooth binary activations by continuation method, in which we begin from learning an easier network with smoothed activation function and let it evolve during the training, until it eventually goes back to being the original, difficult to optimize, deep network with the sign activation function. Comprehensive empirical evidence shows that HashNet can generate exactly binary hash codes and yield state-of-the-art multimedia retrieval performance on standard benchmarks.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/5f824248ff8eb9cd12dd6a13f4ed11a1d696abac

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Deep Multi-View Enhancement Hashing for Image Retrieval (2020, methodology): “Compared with unsupervised hashing algorithm [35] such as LSH [36], KLSH [37] and ITQ [38], supervised methods could obtain more compact binary representaion and generally achieve better retrieval performance [8], [9].”
Deep Cauchy Hashing for Hamming Space Retrieval (2018, methodology): “Previous state-of-theart deep hashing methods, such as DHN [36] and HashNet [4], usually adopt generalized sigmoid function σ (x) = 1/(1 + e−αx) as the probability function.”
MetricGAN+: An Improved Version of MetricGAN for Speech Enhancement (2021, background): “As shown in Figure 2, large (red) behaves like a hard threshold and most output values are either 0 or 1 (more non-smooth and more saturated [26], like a binary mask).”
Two-Stream Deep Hashing With Class-Specific Centers for Supervised Image Search (2020, methodology): “9) HashNet [72] uses a continuation method to address the imbalanced similarity data problem and gradually evolve a tanh function to approximate the sign function.”
The Faiss Library (2024, positive): “Faiss includes a benchmarking framework (bench_fw) that optimizes index types and parameters to explore accuracy , memory usage and search time operating points.”
Pairwise Teacher-Student Network for Semi-Supervised Hashing (2019, methodology): “In ImageNet-100, we use the same data split as HashNet [4] such that select 130 images per class for training, and regard all images in the selected classes from the validation set as the queries.”
Quantization Networks (2019, methodology): “As a specific quantization, HashNet [2] adopts a similar continuous relaxation to train the hash function, where a single tanh function is used for binarization.”
Deep Supervised Hashing for Fast Image Retrieval (2016, methodology): “On the other hand, since NUS-WIDE dataset hasmore similar image pairs than the other two datasets (Cao et al. 2017), andDPSHadopts column sampling strategy over the whole training set in the training stage while our method only computes the loss within a single mini-batch, the difference in ratios h”
Joint-modal Distribution-based Similarity Hashing for Large-scale Unsupervised Deep Cross-modal Retrieval (2020, methodology): “Following [4, 6, 30], we adopt the scaled tanh function starting with a smoothed objective function and becoming more non-smooth with the training process to attack the ill-posed gradient problem.”
Unsupervised Deep Hashing with Similarity-Adaptive and Discrete Optimization (2018, methodology): “[3] propose a new end-to-end deep method dubbed HashNet, which addresses the ill-posed gradient and data imbalance problem.”
Central Similarity Quantization for Efficient Image and Video Retrieval (2019, methodology): “The “deep learning to hash” methods such as CNNH [44], DNNH [14], DHN [48], DCH [1] and Hash-Net [2] have been successfully applied to image hashing.”
Deep Bayesian Hashing with Center Prior for Multi-modal Neuroimage Retrieval (2020, methodology): “Similar to [50], [51], we select the hyper-parameter α via cross-validation.”
Efficient Passage Retrieval with Hashing for Open-domain Question Answering (2021, methodology): “Recently, many deep-learning-to-hash methods have been proposed (Lai et al., 2015; Zhu et al., 2016; Li et al., 2016; Cao et al., 2017, 2018) to jointly learn feature representations and hash functions in an end-to-end manner.”
Binary Generative Adversarial Networks for Image Retrieval (2017, methodology): “In order to tackle this challenge of optimizing deep networks with non-smooth sign activation, we draw inspiration from recent studies on continuation methods [1, 2].”
All the News That Fits in Bits: Learned Rotation-Aware Binary Projections for Efficient News Retrieval at NDTV (2026, unknown): “A parallel line learns binary codes throughsuperviseddeepnetworks,spanningpairwise[16],continuation-based[1],andvariationalorcontrastiveobjectives[2,11]([3,17] forsurveys).”