A Simple Supervised Hashing Algorithm Using Projected Gradient and Oppositional Weights

Abstract

Learning to hash is generating similarity-preserving binary representations of images, which is, among others, an efficient way for fast image retrieval. Two-step hashing has become a common approach because it simplifies the learning by separating binary code inference from hash function training. However, the binary code inference typically leads to an intractable optimization problem with binary constraints. Different relaxation methods, which are generally based on complicated optimization techniques, have been proposed to address this challenge. In this paper, a simple relaxation scheme based on the projected gradient is proposed. To this end in each iteration, we try to update the optimization variable as if there is no binary constraint and then project the updated solution to the feasible set. We formulate the projection step as fining closet binary matrix to the updated matrix and take advantage of the closed-form solution for the projection step to complete our learning algorithm. Inspired by opposition-based learning, pairwise opposite weights between data points are incorporated to impose a stronger penalty on data instances with higher misclassification probability in the proposed objective function. We show that this simple learning algorithm leads to binary codes that achieve competitive results on both CIFAR-10 and NUS-WIDE datasets compared to state-of-the-art benchmarks.

A self-supervised contrastive learning approach for whole slide image representation in digital pathology

“…The results of LUAD/LUSC classification are shown in 3. In this section, we use the features extracted from a DenseNet model, 27 as per Hemati et al 9 We can observe that our suggested strategy outperformed earlier approaches for LUAD/LUSC classification by 2% (delivering 88 %), that underlines the performance of attention-pooling and constrastive learning. We have also, employed the SS-CAMIL blocks in this task and it improved the performance to 89%, but since we are not sure whether it has seen the data in the search task (both datasets are from TCGA repository), we did not report those numbers in the table.…”

“…Table 1 , Table 2 show the horizontal and vertical search results, respectively. We compare our performance with Kalra et al 3 and Hemati et al 9 In both tables, CAMIL is the baseline attention-based MIL with CL and without self-supervision, and SS-CAMIL is the same as CAMIL setup but uses the weights of self-supervision of primary sites (See Table 3 ).…”

“…Again, here we can see that in 15 subtypes, self-supervision has helped the model perform better than CAMIL. To show the effect of CL, we show the 2D t-SNE plot of Hemati et al 9 and SS-CAMIL in Fig. 2 .…”

“… 7 In Adnan et al, 8 the authors used graph CNNs to consider each instance as a node in a graph and then learned an adjacency matrix to build a graph representation of WSIs. Just recently, Hemati et al 9 have exploited deep sets 5 for MIL training in histopathology. They employed a conditional prediction layer where predictions of primary site labels guide the primary diagnosis predictions.…”

On image search in histopathology

“… 23 , 60 Learning tissue representation is a major step toward closing this gap. 21 , 47 Additionally, employing multimodal domain data, such as pathology reports, can help deep learning attach “context” to its otherwise black-box behaviors when it comes to hierarchical tissue representation in their connectionist topologies. 15 , 26

Fig.

…”