The imbalanced distribution of long-tailed data presents a considerable challenge for deep learning models, as it causes them to prioritize the accurate classification head classes but largely disregard tail classes. biased decision boundary caused by inadequate semantic information in is one key factors contributing their low recognition accuracy. To rectify this issue, we propose augment grafting diverse from classes, referred head-to-tail fusion (H2T). We replace portion feature maps with...

Cross-modal hashing (CMH) has attracted considerable attention in recent years. Almost all existing CMH methods primarily focus on reducing the modality gap and semantic gap, i.e., aligning multi-modal features their semantics Hamming space, without taking into account space difference between real number space. In fact, can affect performance of methods. this paper, we analyze demonstrate how affects methods, which therefore raises two problems: solution compression loss function...

It is not uncommon that real-world data are distributed with a long tail. For such data, the learning of deep neural networks becomes challenging because it hard to classify tail classes correctly. In literature, several existing methods have addressed this problem by reducing classifier bias, provided features obtained long-tailed representative enough. However, we find training directly on leads uneven embedding space. That is, space head severely compresses classes, which conducive...

Supervised cross-modal hashing has received wide attention in recent years. However, existing methods primarily rely on sample-wise semantic relationships to evaluate the similarity between samples, overlooking impact of label distribution enhancing retrieval performance. Moreover, limited representation capability traditional dense hash codes hinders preservation relationship. To overcome these challenges, we propose a new method, Joint Semantic Preserving Sparse Hashing (JSPSH)....

As an effective tool for network compression, pruning techniques have been widely used to reduce the large number of parameters in deep neural networks (NNs). Nevertheless, unstructured has limitation dealing with sparse and irregular weights. By contrast, structured can help eliminate this drawback but it requires complex criteria determine which components be pruned. Therefore, paper presents a new method termed BUnit-Net, directly constructs compact NNs by stacking designed basic units,...

Supervised cross-modal hashing methods usually construct a massive undirected weighted graph based on labels for training data, with the aim of learning more structured hash codes by preserving relationships within this graph. However, as volume data increases, such an approach demands substantial computational and storage resources tends to aggregate all points paths, even semantically unrelated ones, which undermines retrieval performance. In paper, we propose prune less crucial paths from...

Knowledge distillation (KD) has become a widely used technique in the field of model compression, which aims to transfer knowledge from large teacher lightweight student for efficient network development. In addition supervision ground truth, vanilla KD method regards predictions as soft labels supervise training model. Based on KD, various approaches have been developed further improve performance However, few these previous methods considered reliability models. Supervision erroneous may...

Convolutional neural networks (CNNs) have achieved significant performance on various real-life tasks. However, the large number of parameters in convolutional layers requires huge storage and computation resources, making it challenging to deploy CNNs memory-constrained embedded devices. In this article, we propose a novel compression method that generates convolution filters each layer using set learnable low-dimensional quantized filter bases. The proposed reconstructs by stacking linear...

Deep Convolutional Neural Networks (CNN) have been successfully applied to many real-life problems. However, the huge memory cost of deep CNN models poses a great challenge deploying them on memory-constrained devices (e.g., mobile phones). One popular way reduce model is train binary where weights in convolution filters are either 1 or -1 and therefore each weight can be efficiently stored using single bit. compression ratio existing upper bounded by ∼ 32. To address this limitation, we...

The imbalanced distribution of long-tailed data presents a considerable challenge for deep learning models, as it causes them to prioritize the accurate classification head classes but largely disregard tail classes. biased decision boundary caused by inadequate semantic information in is one key factors contributing their low recognition accuracy. To rectify this issue, we propose augment grafting diverse from classes, referred head-to-tail fusion (H2T). We replace portion feature maps with...

It is not uncommon that real-world data are distributed with a long tail. For such data, the learning of deep neural networks becomes challenging because it hard to classify tail classes correctly. In literature, several existing methods have addressed this problem by reducing classifier bias provided features obtained long-tailed representative enough. However, we find training directly on leads uneven embedding space. That is, space head severely compresses classes, which conducive...

Lip-password has provided a promising solution for speaker verification (Liu and Cheung 2014). Despite the potential of this technology, there are few related studies, largely attributed to lack corresponding public datasets. Furthermore, previous works in field generally demand substantial amount training samples negative samples, impeding their applications from practical perspective. Therefore, paper collects lip-password dataset proposes novel few-shot based model, which can be...

Deep Convolutional Neural Networks (CNN) have been successfully applied to many real-life problems. However, the huge memory cost of deep CNN models poses a great challenge deploying them on memory-constrained devices (e.g., mobile phones). One popular way reduce model is train binary where weights in convolution filters are either 1 or -1 and therefore each weight can be efficiently stored using single bit. compression ratio existing upper bounded by around 32. To address this limitation,...

Unstructured pruning has the limitation of dealing with sparse and irregular weights. By contrast, structured can help eliminate this drawback but it requires complex criterion to determine which components be pruned. To end, paper presents a new method termed TissueNet, directly constructs compact neural networks fewer weight parameters by independently stacking designed basic units, without requiring additional judgement criteria anymore. Given units various architectures, they are...

Existing convolutional neural networks (CNNs) have achieved significant performance on various real-life tasks, but a large number of parameters in layers requires huge storage and computation resources which makes it difficult to deploy CNNs memory-constraint embedded devices. In this paper, we propose novel compression method that generates the convolution filters each layer by combining set learnable low-dimensional binary filter bases. The proposed designs more compact stacking linear...

Existing convolutional neural networks (CNNs) have achieved significant performance on various real-life tasks, but a large number of parameters in layers requires huge storage and computation resources which makes it difficult to deploy CNNs memory-constraint embedded devices. In this paper, we propose novel compression method that generates the convolution filters each layer by combining set learnable low-dimensional binary filter bases. The proposed designs more compact stacking linear...