Recently, Vision Transformers (ViTs) have become the state-of-the-art architecture on various computer vision tasks including image classification, object detection and semantic segmentation. However, such success in high-accuracy performance comes at the price of high computational complexity, with typically tens of millions of or even more parameters in a Vision Transformer (ViT) model. Such a large volume of parameters makes it very difficult to deploy ViT models on mobile devices and cumbers their applications. In this paper, we present a novel post-training quantization approach that is able to quantize ViT models to very low bit widths, without the need of re-training. Prior works on post-training quantization for ViTs optimize the quantization of each layer separately thus leading to sub-optimal results. In contrast, we propose a unified learning framework that jointly optimizes the quantization of all layers to directly reduce the overall output error of the network. Moreover, we explore an important property of ViTs, i.e., the additivity property, revealing that the output error caused by the quantization of multiple layers equals the sum of the output error due to the quantization of each layer. Utilizing this property, we present a very efficient algorithm to solve the joint optimization problem with linear time complexity. We performed extensive experiments on the large-scale ImageNet dataset to evaluate the effectiveness of our approach. Empirical results show that our approach improves state-of-the-art noticeably on various ViT models and lowers the bit width from 8-bit to 6-bit without hurting the accuracy. Specifically, at 4 bits, our approach significantly outperforms existing works by 1.72%, 11.49%, 6.15%, and 3.54% on ViT-S, ViT-B, DeiT-S, and DeiT-B, respectively. In the end, we evaluate the performance when deploying our quantized models on hardware. Our approach achieves $1.5\times $ to $1.7\times $ speedups for the inference on NVIDIA A100 GPU.

In this paper, we propose a post-training pruning framework that jointly optimizes layerwise pruning to minimize model output distortion. Through theoretical and empirical analysis, we discover an important additivity property of output distortion from pruning weights/channels in DNNs. Leveraging this property, we reformulate pruning optimization as a combinatorial problem and solve it with dynamic programming, achieving linear time complexity and making the algorithm very fast on CPUs. Furthermore, we optimize additivity-derived distortions using Hessian-based Taylor approximation to enhance pruning efficiency, accompanied by fine-grained complexity reduction techniques. Our method is evaluated on various DNN architectures, including CNNs, ViTs, and object detectors, and on vision tasks such as image classification on CIFAR-10 and ImageNet, and 3D object detection and various datasets. We achieve SoTA with significant FLOPs reductions without accuracy loss. Specifically, on CIFAR-10, we achieve up to $27.9\times$27.9×, $29.2\times$29.2×, and $14.9\times$14.9× FLOPs reductions on ResNet-32, VGG-16, and DenseNet-121, respectively. On ImageNet, we observe no accuracy loss with $1.69\times$1.69× and $2\times$2× FLOPs reductions on ResNet-50 and DeiT-Base, respectively. For 3D object detection, we achieve $\mathbf {3.89}\times, \mathbf {3.72}\times$3.89×,3.72× FLOPs reductions on CenterPoint and PVRCNN models. These results demonstrate the effectiveness and practicality of our approach for improving model performance through layer-adaptive weight pruning.

Deep neural networks (DNNs) have been widely used in many artificial intelligence (AI) tasks. However, deploying them brings significant challenges due to the huge cost of memory, energy, and computation. To address these challenges, researchers have developed various model compression techniques such as model quantization and model pruning. Recently, there has been a surge in research on compression methods to achieve model efficiency while retaining performance. Furthermore, more and more works focus on customizing the DNN hardware accelerators to better leverage the model compression techniques. In addition to efficiency, preserving security and privacy is critical for deploying DNNs. However, the vast and diverse body of related works can be overwhelming. This inspires us to conduct a comprehensive survey on recent research toward the goal of high-performance, cost-efficient, and safe deployment of DNNs. Our survey first covers the mainstream model compression techniques, such as model quantization, model pruning, knowledge distillation, and optimizations of nonlinear operations. We then introduce recent advances in designing hardware accelerators that can adapt to efficient model compression approaches. In addition, we discuss how homomorphic encryption can be integrated to secure DNN deployment. Finally, we discuss several issues, such as hardware evaluation, generalization, and integration of various compression approaches. Overall, we aim to provide a big picture of efficient DNNs from algorithm to hardware accelerators and security perspectives.

In this paper, we propose a novel layer-adaptive weight-pruning approach for Deep Neural Networks (DNNs) that addresses the challenge of optimizing the output distortion minimization while adhering to a target pruning ratio constraint. Our approach takes into account the collective influence of all layers to design a layer-adaptive pruning scheme. We discover and utilize a very important additivity property of output distortion caused by pruning weights on multiple layers. This property enables us to formulate the pruning as a combinatorial optimization problem and efficiently solve it through dynamic programming. By decomposing the problem into sub-problems, we achieve linear time complexity, making our optimization algorithm fast and feasible to run on CPUs. Our extensive experiments demonstrate the superiority of our approach over existing methods on the ImageNet and CIFAR-10 datasets. On CIFAR-10, our method achieves remarkable improvements, outperforming others by up to 1.0% for ResNet-32, 0.5% for VGG-16, and 0.7% for DenseNet-121 in terms of top-1 accuracy. On ImageNet, we achieve up to 4.7% and 4.6% higher top-1 accuracy compared to other methods for VGG-16 and ResNet-50, respectively. These results highlight the effectiveness and practicality of our approach for enhancing DNN performance through layer-adaptive weight pruning. Code will be available on https://github.com/Akimoto-Cris/RD_VIT_PRUNE.

A popular track of network compression approach is Quantization aware Training (QAT), which accelerates the forward pass during the neural network training and inference. However, not much prior efforts have been made to quantize and accelerate the backward pass during training, even though that contributes around half of the training time. This can be partly attributed to the fact that errors of low-precision gradients during backward cannot be amortized by the training objective as in the QAT setting. In this work, we propose to solve this problem by incorporating the gradients into the computation graph of the next training iteration via a hypernetwork. Various experiments on CIFAR-10 dataset with different CNN network architectures demonstrate that our hypernetwork-based approach can effectively reduce the negative effect of gradient quantization noise and successfully quantizes the gradients to INT4 with only 0.64 accuracy drop for VGG-16 on CIFAR-10.

As a successor to convolutional neural networks (CNNs), transformer-based models have achieved great performance in computer vision tasks. Compressing vision transformers to low-bit brings a number of practical benefits, including higher inference speed, improved memory footprint, and reduced energy consumption. Existing model compression methods, especially quantization techniques, ignore the joint statistics of weights, resulting in sub-optimal task performance at a given quantization bit rate. In this paper, we propose to apply a transform before quantization to decorrelate vision transformer’s weights. And the entire compression flow is optimized in a rate-distortion framework to minimize the network output errors instead of simply optimizing for quantization errors or layer-wise output errors. Extensive experimental results on a variety of vision transformers (e.g. Swin, ViT and DeiT) demonstrate that our proposed method outperforms the state-of-the-art. It can quantize vision transformers (e.g. Swin, ViT and DeiT) on both weights and activations to 6-bit without a significant accuracy drop.

Intermediate deep visual feature compression and transmission is an emerging research topic, which enables a good balance among computing load, bandwidth usage and generalization ability for AI-based visual analysis in edge-cloud collaboration. Quantization and the corresponding rate-distortion optimization are the key techniques in deep feature compression. In this paper, by exploring the feature statistics and a greedy iterative algorithm, we propose a channel-wise bit allocation method for deep feature quantization optimizing for network output error. Given the limited rate and computational power, the proposed method can quantize features with small information loss. Moreover, the method also provides the option to handle the trade-offs between computational cost and quantization performance. Experimental results on ResNet and VGGNet features demonstrate the effectiveness of the proposed bit allocation method.

In this paper, we compress convolutional neural network (CNN) weights post-training via transform quantization. Previous CNN quantization techniques tend to ignore the joint statistics of weights and activations, producing sub-optimal CNN performance at a given quantization bit-rate, or consider their joint statistics during training only and do not facilitate efficient compression of already trained CNN models. We optimally transform (decorrelate) and quantize the weights post-training using a rate-distortion framework to improve compression at any given quantization bit-rate. Transform quantization unifies quantization and dimensionality reduction (decorrelation) techniques in a single framework to facilitate low bit-rate compression of CNNs and efficient inference in the transform domain. We first introduce a theory of rate and distortion for CNN quantization, and pose optimum quantization as a rate-distortion optimization problem. We then show that this problem can be solved using optimal bit-depth allocation following decorrelation by the optimal End-to-end Learned Transform (ELT) we derive in this paper. Experiments demonstrate that transform quantization advances the state of the art in CNN compression in both retrained and non-retrained quantization scenarios. In particular, we find that transform quantization with retraining is able to compress CNN models such as AlexNet, ResNet and DenseNet to very low bit-rates (1-2 bits).

Heterogeneous information networks (HINs) are usually used to model information systems with multi-type objects and relations. In contrast, graphs that have a single type of nodes and edges, are often called homogeneous graphs. Measuring similarities among objects is an important task in data mining applications, such as web search, link prediction, and clustering. Currently, several similarity measures are defined for HINs. Most of these measures are based on meta-paths, which show sequences of node classes and edge types along the paths between two nodes. However, meta-paths, which are often designed by domain experts, are hard to enumerate and choose w.r.t. the quality of similarity scores. This makes using existing similarity measures in real applications difficult. To address this problem, we extend SimRank, a well-known similarity measure on homogeneous graphs, to HINs, by introducing the concept of the decay graph. The newly proposed similarity measure is called HowSim, which has the property of being meta-path free, and capturing the structural and semantic similarity simultaneously. The generality and effectiveness of HowSim, and the efficiency of our proposed algorithms for computing HowSim scores, are demonstrated by extensive experiments.

Power and area-efficient deep neural network (DNN) designs are key in edge applications. Compact DNNs, via compression or quantization, enable such designs by significantly reducing memory footprint. Lossless entropy coding can further reduce the size of networks. It is then critical to provide hardware support for such entropy coding module to fully benefit from the resulted reduced memory requirement. In this work, we introduce Tunstall coding to compress the quantized weights. Tunstall coding can achieve high compression ratio as well as very fast decoding speed on various deep networks. We present two hardware-accelerated decoding techniques that provide streamlined decoding capabilities. We synthesize these designs targeting on FPGA. Results show that we achieve up to 6× faster decoding time versus state-of-the-art decoding methods.

Learning network embeddings has attracted growing attention in recent years. However, most of the existing methods focus on homogeneous networks, which cannot capture the important type information in heterogeneous networks. To address this problem, in this paper, we propose TransN, a novel multi-view network embedding framework for heterogeneous networks. Compared with the existing methods, TransN is an unsupervised framework which does not require node labels or user-specified meta-paths as inputs. In addition, TransN is capable of handling more general types of heterogeneous networks than the previous works. Specifically, in our framework TransN, we propose a novel algorithm to capture the proximity information inside each single view. Moreover, to transfer the learned information across views, we propose an algorithm to translate the node embeddings between different views based on the dual-learning mechanism, which can both capture the complex relations between node embeddings in different views, and preserve the proximity information inside each view during the translation. We conduct extensive experiments on real-world heterogeneous networks, whose results demonstrate that the node embeddings generated by TransN outperform those of competitors in various network mining tasks.

Heterogeneous information networks (HINs) are usually used to model information systems with multi-type objects and relations. Measuring the similarity among objects is an important task in data mining applications. Currently, several similarity measures are defined for HIN. Most of these measures are based on meta-paths, which show sequences of node classes and edge types along the paths between two nodes. However, meta-paths, which are often designed by domain experts, are hard to enumerate and choose w.r.t. the quality of the similarity scores. This makes the existing similarity measures difficult to use in real applications. To address this problem, we extend SimRank, a well-known similarity measure for homogeneous graphs, to HINs, by introducing the concept of decay graph. The newly proposed relevance measure is called HowSim, which has the property of being meta-path free, and capturing the structural and semantic similarity simultaneously. The generality and effectiveness of HowSim, are demonstrated by extensive experiments.

Modern convolutional object detectors have improved the detection accuracy significantly, which in turn inspired the development of dedicated hardware accelerators to achieve real-time performance by exploiting inherent parallelism in the algorithm. Non-maximum suppression (NMS) is an indispensable operation in object detection. In stark contrast to most operations, the commonly-adopted GreedyNMS algorithm does not foster parallelism, which can be a major performance bottleneck. In this paper, we introduce MaxpoolNMS, a parallelizable alternative to the NMS algorithm, which is based on max-pooling classification score maps. By employing a novel multi-scale multi-channel max-pooling strategy, our method is 20x faster than GreedyNMS while simultaneously achieves comparable accuracy, when quantified across various benchmarking datasets, i.e., MS COCO, KITTI and PASCAL VOC. Furthermore, our method is better suited for hardware-based acceleration than GreedyNMS.

In this paper, we propose Deep Holographic Networks (DHN) to learn similarity metrics of videos for multi-label video search. DHN introduces a holographic composition layer to explicitly encode similarity metrics at intermediate layer of the network, instead of conventional deep metric learning approaches driven by ranking losses. The holographic composition layer is parameter-free and enables less memory footprint compared with state-of-the-art. Towards multi-label video search at large scale, we present a new video benchmark built upon the YouTube-8M dataset. Extensive evaluations on this dataset demonstrate that DHN performs better than traditional deep metric learning approaches as well as other compositional networks.

Hashing, a widely-studied solution to the approximate nearest neighbor (ANN) search, aims to map data points in the high-dimensional Euclidean space to the low-dimensional Hamming space while preserving the similarity between original points. As directly learning binary codes can be NP-hard due to discrete constraints, a two-stage scheme, namely "projection and quantization", has already become a standard paradigm for learning similarity-preserving hash codes. However, most existing hashing methods typically separate these two stages and thus fail to investigate complementary effects of both stages. In this paper, we systematically study the relationship between "projection and quantization", and propose a novel minimal reconstruction bias hashing (MRH) method to learn compact binary codes, in which the projection learning and quantization optimizing are jointly performed. By introducing a lower bound analysis, we design an effective ternary search algorithm to solve the corresponding optimization problem. Furthermore, we conduct some insightful discussions on the proposed MRH approach, including the theoretical proof, and computational complexity. Distinct from previous works, MRH can adaptively adjust the projection dimensionality to balance the information loss between projection and quantization. The proposed framework not only provides a unique perspective to view traditional hashing methods but also evokes some other researches, e.g., guiding the design of the loss functions in deep networks. Extensive experiment results have shown that the proposed MRH significantly outperforms a variety of state-of-the-art methods over eight widely used benchmarks.

This paper proposes a novel algorithm to solve the exact r-neighbour search problem in Hamming space. Existing r-neighbour search methods typically adopt hash table to index binary codes. Given a query, existing approaches search the nearest neighbours by checking all buckets of a Hamming ball centered at the query. The problem is these methods spend most of search time visiting empty buckets (lookup misses). In this paper, we adopt trie structure to index binary codes. We consider several continuous bits of a binary string as a block and use it as an atomic indexing element in trie structure, which is efficient in access speed and memory usage. Our method searches the nearest neighbours of a query by utilizing the records of nodes in trie structure to avoid lookup misses. We name the proposed indexing structure as Multi-Block N-ary Trie (MBNT). A theoretical analysis is given to prove that MBNT has less time cost than other hash-based methods. Extensive results show that MBNT outperforms state-of-the-art algorithms on several large scale benchmarks.

Hashing techniques are powerful for approximate nearest neighbour (ANN) search.Existing quantization methods in hashing are all focused on scalar quantization (SQ) which is inferior in utilizing the inherent data distribution.In this paper, we propose a novel vector quantization (VQ) method named affinity preserving quantization (APQ) to improve the quantization quality of projection values, which has significantly boosted the performance of state-of-the-art hashing techniques.In particular, our method incorporates the neighbourhood structure in the pre- and post-projection data space into vector quantization.APQ minimizes the quantization errors of projection values as well as the loss of affinity property of original space.An effective algorithm has been proposed to solve the joint optimization problem in APQ, and the extension to larger binary codes has been resolved by applying product quantization to APQ.Extensive experiments have shown that APQ consistently outperforms the state-of-the-art quantization methods, and has significantly improved the performance of various hashing techniques.

To achieve high retrieval accuracy over a large scale image/video dataset, recent research efforts have demonstrated that employing extremely high-dimensional descriptors such as the Fisher Vector (FV) and the Vector of Locally Aggregated Descriptors (VLAD) can yield good performance. To enable fast search, the FV (or VLAD) is usually compressed by product quantization (PQ) or hashing. However, compressing high-dimensional descriptors via PQ or hashing may become intractable and infeasible due to both the storage and computation requirements for the linear/nonlinear projection of PQ or hashing methods. We develop a novel compact aggregated descriptor via dual selection for visual search. We utilize both sample-specific Gaussian component redundancy and bit dependency within a binary aggregated descriptor to produce its compact binary codes. The proposed method can effectively reduce the codesize of the raw aggregated descriptors, without degrading the search accuracy or introducing additional memory footprint. We demonstrate the significant advantages of the proposed binary codes in solving the approximate nearest neighbor (ANN) visual search problem. Experimental results on extensive datasets show that our method outperforms the state-of-the-art methods.

Compact descriptors for visual search (CDVS) is a recently completed standard from the ISO/IEC moving pictures experts group (MPEG). The primary goal of this standard is to provide a standardized bitstream syntax to enable interoperability in the context of image retrieval applications. Over the course of the standardization process, remarkable improvements were achieved in reducing the size of image feature data and in reducing the computation and memory footprint in the feature extraction process. This paper provides an overview of the technical features of the MPEG-CDVS standard and summarizes its evolution.

Fisher vectors (FV) aggregated from local invariant features (e.g., SIFT) is one of the state-of-the-art descriptors for visual search, due to high discriminability but small visual vocabulary. Nevertheless, a high-dimensional FV needs to be compressed into a compact descriptor for light storage and high matching eficiency. In this paper, we formulate the FV compression as a resource-constrained optimization problem. Our goal is to maximize search performance subject to the constraints of descriptor compactness, compression complexity in terms of memory usage and time cost. Accordingly, we present a selective binary Fisher codes (SBFC) to compress the raw FV. Firstly, to fulfill the constraint of compression complexity, we binarize the FV by a sign function, Secondly, we propose to select discriminative bits from the binarized FV (BFC) to maximize search performance, subject to the constraint of descriptor compactness. Extensive experiments over MPEG Compact Descriptor for Visual Search (CDVS) benchmark datasets have shown that S-BFC significantly improves search performance at a smaller descriptor size as well as much lower complexity, compared with the state-of-the-art FV compression algorithms like Hashing and Product Quantziation (PQ). A simplified version of SBFC, SBFC LS has been adopted by the MPEG CDVS standard. In the CDVS evaluation framework, SBFC LS has achieved promising performance mean Average Precision (mAP) 83% on average at much lower memory cost of 40KB.

Constructing discriminative feature descriptors is crucial towards effective image retrieval. The state-of-the-art powerful global descriptor for this purpose is Vector of Locally Aggregated Descriptors (VLAD). Given a set of local features (say, SIFT) extracted from an image, the VLAD is generated by quantizing local features with a small visual vocabulary (64 to 512 centroids), aggregating the residual statistics of quantized features for each centroid and concatenating the aggregated residual vectors from each centroid. One can increase the search accuracy by increasing the size of vocabulary (from hundreds to hundreds of thousands), which, however, it leads to heavy computation cost with flat quantization. In this paper, we propose a hierarchical multi-VLAD to seek the tradeoff between descriptor discriminability and computation complexity. We build up a tree-structured hierarchical quantization (TSHQ) to accelerate the VLAD computation with a large vocabulary. As quantization error may propagate from root to leaf node (centroid) with TSHQ, we introduce multi-VLAD, which constructing a VLAD descriptor for each level of the vocabulary tree, so as to compensate for the quantization error at that level. Extensive evaluation over benchmark datasets has shown that the proposed approach outperforms state-of-the-art in terms of retrieval accuracy, fast extraction, as well as light memory cost.

Towards low bit rate mobile visual search, recent works have proposed to aggregate the local features and compress the aggregated descriptor (such as Fisher vector, the vector of locally aggregated descriptors) for low latency query delivery as well as moderate search complexity. Even though Hamming distance can be computed very fast, the computational cost of exhaustive linear search over the binary descriptors grows linearly with either the length of a binary descriptor or the number of database images. In this paper, we propose a novel weighted component hashing (WeCoHash) algorithm for long binary aggregated descriptors to significantly improve search efficiency over a large scale image database. Accordingly, the proposed WeCoHash has attempted to address two essential issues in Hashing algorithms: “what to hash” and “how to search.” “What to hash” is tackled by a hybrid approach, which utilizes both image-specific component (i.e., visual word) redundancy and bit dependency within each component of a binary aggregated descriptor to produce discriminative hash values for bucketing. “How to search” is tackled by an adaptive relevance weighting based on the statistics of hash values. Extensive comparison results have shown that WeCoHash is at least 20 times faster than linear search and 10 times faster than local sensitive hash (LSH) when maintaining comparable search accuracy. In particular , the WeCoHash solution has been adopted by the emerging MPEG compact descriptor for visual search (CDVS) standard to significantly speed up the exhaustive search of the binary aggregated descriptors.

Compact locally aggregated binary features have shown great advantages in image search. As the exhaustive linear search in Hamming space still entails too much computational complexity for large datasets, recent works proposed to directly use binary codes as hash indices, yielding a dramatic increase in speedup. However, these methods cannot be directly applied to variable-length binary features. In this paper, we propose a Component Hashing (CoHash) algorithm to handle the variable-length binary aggregated descriptors indexing for fast image search. The main idea is to decompose the distance measure between variable-length descriptors into aligned component-to-component matching problems independently, and build multiple hash tables for the visual word components. Given a query, its candidate neighbors are found by using the query binary sub-vectors as indices into their corresponding hash tables. In particular, a bit selection based on conditional mutual information maximization is proposed to reduce the dimensionality of visual word components, which provides a light storage of indices and balances the retrieval accuracy and search cost. Extensive experiments on benchmark datasets show that our approach is 20~25 times faster than linear search, without any noticeable retrieval performance loss.

While much progress has been made in mobile visual search, user experiences still relate to the query transmission latency, especially over a bandwidth-constrained wireless link. Low bit rate visual search paradigm has been well advocated in both academic and industrial endeavors, which directly extracts and sends compact visual descriptor(s) rather than sending a query image. Recent advances in compact descriptor design have advocated the use of compressed bag-of-words histogram, which has shown superior performance over other alternatives. However, existing works focus on descriptor compactness, regardless of time cost and memory requirements on the extraction pipeline, which in turn is crucial for the mobile end development. In this paper, we investigate the problem of designing a memory-light descriptor extraction scheme based upon the so-called multi-stage vector quantization. Our scheme starts by quantizing local patches with a small codebook, and the resulting quantization residual is subsequently compensated by a product quantizer. The design of both quantizers are based upon improving PSNR, which would drop a lot through quantization. PSNR is quantitatively shown to be highly correlated with retrieval and matching accuracy. Extensive evaluation on MPEG Compact Descriptor for Visual Search (CDVS) dataset, has reported superior performance over the state-of-the-art.