Recent years have witnessed growing scholarly interest in binary post-training quantization (PTQ) techniques for large language models (LLMs). While state-of-the-art (SOTA) binary quantization methods significantly reduce memory footprint and computational demands, they introduce additional memory overhead beyond binary weight tensors to mitigate performance degradation. Moreover, binary LLMs still suffer from substantial accuracy loss. To address these limitations, we propose MemeBQ, a novel binary PTQ framework for LLMs that reduces the memory overhead of auxiliary flag bitmaps in existing binary quantization methods. Specifically, we first design a greedy row clustering method, which leverages the similarity between the row vectors of weights to partition the weight rows into different groups. By sharing the common flag bitmap within each row group, we significantly mitigate the memory overhead associated with flag bitmaps. Besides, to improve the performance of binary LLMs, we propose a novel weight splitting method for each row group of weights, which determines the flag bitmap's values in a fine-grained way. Extensive experiments on OPT, Llama-2, and Llama-3 models demonstrate that MemeBQ reduces 50% extra memory demand while achieving comparable accuracy compared with current SOTA methods. Alternatively, MemeBQ outperforms SOTA binary quantization methods up to 7% with the same extra bits on reasoning benchmarks.

Low-bit quantization of Vision Transformers presents significant challenges due to the intrinsic properties of their Multi-head Self-attention modules. In this work, we investigate the quantization issues specific to this mechanism and identify two critical challenges. First, quantization amplifies discrepancies among attention heads, thereby impairing the model’s capability to focus on the most informative regions. Second, high-magnitude values in the attention maps, which encode essential relational information, exhibit an extremely sharp distribution that renders them especially prone to substantial information loss during quantization. To address these challenges, we propose Quantized-aware Multi-head Self-attention (Q-MHSA), a universal self-attention enhancement module that integrates two lightweight components within Multi-head Self-attention (MHSA). The Cross-head Concordance Module (CCM) enforces adaptive consistency across attention heads, while the Learnable Smoothness Controller (LSC) replaces the fixed normalization factor with an adaptive mechanism that selectively smooths the distribution of high-magnitude attention values while disregarding less informative low values. Designed for seamless integration with any Vision Transformer architecture and quantization-aware training method, Q-MHSA incurs minimal overhead while consistently improving model accuracy. For instance, under the VVTQ quantization framework applied to a 4-bit Swin-S model, the incorporation of Q-MHSA yields a top-1 accuracy of 83.5%, representing a 0.9% improvement over the baseline, while incurring a marginal overhead of 0.01% in both parameters and FLOPs.

Multi-bit spiking neural networks (SNNs) have recently become a heated research spot, pursuing energy-efficient and high-accurate AI. However, with more bits involved, the associated memory and computation demands escalate to the point where the performance improvements become disproportionate. Based on the insight that different layers demonstrate different importance and extra bits could be wasted and interfering, this paper presents an adaptive bit allocation strategy for direct-trained SNNs, achieving fine-grained layer-wise allocation of memory and computation resources. Thus, SNN's efficiency and accuracy can be improved. Specifically, we parametrize the temporal lengths and the bit widths of weights and spikes, and make them learnable and controllable through gradients. To address the challenges caused by changeable bit widths and temporal lengths, we propose the refined spiking neuron, which can handle different temporal lengths, enable the derivation of gradients for temporal lengths, and suit spike quantization better. In addition, we theoretically formulate the step-size mismatch problem of learnable bit widths, which may incur severe quantization errors to SNN, and accordingly propose the step-size renewal mechanism to alleviate this issue. Experiments on various datasets, including the static CIFAR and ImageNet datasets and the dynamic CIFAR-DVS and DVS-GESTURE datasets, demonstrate that our methods can reduce the overall memory and computation cost while achieving higher accuracy. Particularly, our SEWResNet-34 can achieve a 2.69 % accuracy gain and 4.16 × lower bit budgets over the advanced baseline work on ImageNet. This work is open-sourced at this link.

Post-Training Quantization is an effective technique for deep neural network acceleration. However, as the bit-width decreases to 4 bits and below, PTQ faces significant challenges in preserving accuracy, especially for attention-based models like LLMs. The main issue lies in considerable clipping and rounding errors induced by the limited number of quantization levels and narrow data range in conventional low-precision quantization. In this paper, we present an efficient and accurate PTQ method that targets 4 bits and below through algorithm and architecture co-design. Our key idea is to dynamically extract a small portion of significant bit terms from high-precision operands to perform low-precision multiplications under the given computational budget. Specifically, we propose Significant-Bit Quantization (SBQ). It exploits a product-aware method to dynamically identify significant terms and an error-compensated computation scheme to minimize compute errors. We present a dedicated inference engine to unleash the power of SBQ. Experiments on CNNs, ViTs, and LLMs reveal that SBQ consistently outperforms prior PTQ methods under 2~4-bit quantization. We also compare the proposed inference engine with state-of-the-art bit-operation-based quantization architectures TQ and Sibia. Results show that SBQ can achieve the highest area and energy efficiency.

Graph Neural Networks (GNNs) have shown great superiority on non-Euclidean graph data, achieving ground-breaking performance on various graph-related tasks. As a practical solution to train GNN on large graphs with billions of nodes and edges, the sampling-based training is widely adopted by existing training frameworks. However, through an in-depth analysis, we observe that the efficiency of existing sampling-based training frameworks is still limited due to the key bottlenecks lying in all three phases of sampling-based training, i.e., subgraph sample, memory IO, and computation. To this end, we propose FastGL, a GPU-efficient Framework for accelerating sampling-based training of GNN at Large scale by simultaneously optimizing all above three phases, taking into account both GPU characteristics and graph structure. Specifically, by exploiting the inherent overlap within graph structures, FastGL develops the Match-Reorder strategy to reduce the data traffic, which accelerates the memory IO without incurring any GPU memory overhead. Additionally, FastGL leverages a Memory-Aware computation method, harnessing the GPU memory's hierarchical nature to mitigate irregular data access during computation. FastGL further incorporates the Fused-Map approach aimed at diminishing the synchronization overhead during sampling. Extensive experiments demonstrate that FastGL can achieve an average speedup of 11.8×, 2.2× and 1.5× over the state-of-the-art frameworks PyG, DGL, and GNNLab, respectively. Our code is available at https://github.com/a1bc2def6g/fastgl-ae.

Graph Neural Networks (GNNs) are becoming a promising technique in various domains due to their excellent capabilities in modeling non-Euclidean data. Although a spectrum of accelerators has been proposed to accelerate the inference of GNNs, our analysis demonstrates that the latency and energy consumption induced by DRAM access still significantly impedes the improvement of performance and energy efficiency. To address this issue, we propose a Memory - Efficient GNN Accelerator (MEGA) through algorithm and hardware co-design in this work. Specifically, at the algorithm level, through an in-depth analysis of the node property, we observe that the data-independent quantization in previous works is not optimal in terms of accuracy and memory efficiency. This motivates us to propose the Degree-Aware mixed-precision quantization method, in which a proper bitwidth is learned and allocated to a node according to its in-degree to compress GNNs as much as possible while maintaining accuracy. At the hardware level, we employ a heterogeneous architecture design in which the aggregation and combination phases are implemented separately with different dataflows. In order to boost the performance and energy efficiency, we also present an Adaptive-Package format to alleviate the storage overhead caused by the fine-grained bitwidth and diverse sparsity, and a Condense-Edge scheduling method to enhance the data locality and further alleviate the access irregularity induced by the extremely sparse adjacency matrix in the graph. We implement our MEGA accelerator in a 28nm technology node. Extensive experiments demonstrate that MEGA can achieve an average speedup of 38.3 ×, 7.1 ×, 4.0 ×, 3.6× and 47.6 ×, 7.2 ×, 5.4 ×, 4.5 × energy savings over four state-of-the-art GNN accelerators, HyGCN, GCNAX, GROW, and SGCN, respectively, while retaining task accuracy.

Heatmap-based methods have dominated the face alignment task, yet the maximum response decoding scheme necessitates further reform. While some studies have attempted to compensate for prediction offsets using a post-processing module, the prediction errors induced by the maximum response decoding scheme remain challenging to rectify. In this paper, we assume that using heatmap value to denote the ground-truth probability is not accurate enough. To cure this problem, we propose DISPAL, a novel DIStribution-based Probability for fAcial Landmarks, which signifies the ground-truth probability by the similarity between the pixel’s neighbouring value distribution and Gaussian distribution. This innovative probability enables us to pinpoint the keypoint location more robustly than previous methods that rely solely on the peak score. It also exhibits remarkable generalization to complex decoding methodologies. Furthermore, we propose supervising this probability as an additional task loss to help the model learn better heatmap representation. Extensive empirical results on WFLW, 300W, and COFW datasets demonstrate that our distribution-based probability mechanism significantly surpasses original value-based probability approaches.

Neural Radiance Fields (NeRFs) have revolutionized 3D scene synthesis. Voxel grids are commonly employed to enhance training or rendering speed, but they entail additional storage requirements. The large model size and high computational and memory demands impede their progress on resource-constrained devices, e.g., Microcontroller Units (MCUs). Besides, there is currently no NeRF rendering framework available on MCU devices. In this paper, we propose a NeRF method named MCUNeRF for 3D scene synthesis on MCU devices. The proposed MCUNeRF compresses voxel grids via a hybrid quantization algorithm merging learned step-size quantization (LSQ) and optimized product quantization (OPQ). To further reduce the model storage, we also propose a codebook-sharing method that renders multiple objects with a single quantization codebook. Then we implement a NeRF-based rendering framework for MCU devices, which leverages a low-bit neural network computation framework, i.e. CMSIS-NN, to accelerate the rendering progress. Extensive experiments on four datasets such as Synthetic-NeRF demonstrate that our proposed method could compress model data by 20-40 times with comparable rendering quality, which enables NeRF-based scene rendering on MCU devices with only 1M SRAM.

In recent years, various face-landmark datasets have been published. Intuitively, it is significant to integrate multiple labeled datasets to achieve higher performance. Due to the different annotation schemes of datasets, it is hard to directly train models using them together. Although numerous efforts have been made in the joint use of datasets, there remain three shortages in previous methods,i.e., additional computation, limitation of the markups scheme, and limited support for the regression method. To solve the above issues, we proposed a novelAlternating Training Framework(ATF), which leverages the similarity and diversity across multiple datasets for a more robust detector. ATF mainly contains two sub-modules:Alternating Training with Decreasing Proportions(ATDP) andMixed Branch Loss($\mathcal {L}_{MB}$). In particular, ATDP trains multiple datasets simultaneously via a weakly supervised way to take advantage of the diversity among them, and$\mathcal {L}_{MB}$utilizes similar landmark pairs to constrain different branches of the corresponding datasets. Besides, we extend the framework to easily handle three situations: single target detector, joint detector, and novel detector. Extensive experiments demonstrate the effectiveness of our framework for both heatmap-based and direct coordinate regression. Moreover, we have achieved a joint detector that outperforms state-of-the-art methods on each benchmark.

Direct convolution methods are now drawing increasing attention as they eliminate the additional storage demand required by indirect convolution algorithms (i.e., the transformed matrix generated by the im2col convolution algorithm). Nevertheless, the direct methods require special input-output tensor formatting, leading to extra time and memory consumption to get the desired data layout. In this article, we show that indirect convolution, if implemented properly, is able to achieve high computation performance with the help of highly optimized subroutines in matrix multiplication while avoid incurring substantial memory overhead. The proposed algorithm is called efficient convolution via blocked columnizing (ECBC). Inspired by the im2col convolution algorithm and the block algorithm of general matrix-to-matrix multiplication, we propose to conduct the convolution computation blockwisely. As a result, the tensor-to-matrix transformation process (e.g., the im2col operation) can also be done in a blockwise manner so that it only requires a small block of memory as small as the data block. Extensive experiments on various platforms and networks validate the effectiveness of ECBC, as well as the superiority of our proposed method against a set of widely used industrial-level convolution algorithms.

While local-window self-attention performs notably in vision tasks, it suffers from limited receptive field and weak modeling capability issues. This is mainly because it performs self-attention within non-overlapped windows and shares weights on the channel dimension. We propose Mix-Former to find a solution. First, we combine local-window self-attention with depth-wise convolution in a parallel design, modeling cross-window connections to enlarge the receptive fields. Second, we propose bi-directional interactions across branches to provide complementary clues in the channel and spatial dimensions. These two designs are integrated to achieve efficient feature mixing among windows and dimensions. Our MixFormer provides competitive results on image classification with EfficientNet and shows better results than RegNet and Swin Transformer. Performance in downstream tasks outperforms its alternatives by significant margins with less computational costs in 5 dense prediction tasks on MS COCO, ADE20k, and LVIS. Code is available at https://github.com/PaddlePaddle/PaddleClas.

Hashing is a popular search algorithm for its compact binary representation and efficient Hamming distance calculation. Benefited from the advance of deep learning, deep hashing methods have achieved promising performance. However, those methods usually learn with expensive labeled data but fail to utilize unlabeled data. Furthermore, the traditional pairwise loss used by those methods cannot explicitly force similar/dissimilar pairs to small/large distances. Both weaknesses limit existing methods’ performance. To solve the first problem, we propose a novel semi-supervised deep hashing model named adversarial binary mutual learning (ABML). Specifically, our ABML consists of a generative model$G_{H}$and a discriminative model$D_{H}$, where$D_{H}$learns labeled data in a supervised way and$G_{H}$learns unlabeled data by synthesizing real images. We adopt an adversarial learning (AL) strategy to transfer the knowledge of unlabeled data to$D_{H}$by making$G_{H}$and$D_{H}$mutually learn from each other. To solve the second problem, we propose a novel Weibull cross-entropy loss (WCE) by using the Weibull distribution, which can distinguish tiny differences of distances and explicitly force similar/dissimilar distances as small/large as possible. Thus, the learned features are more discriminative. Finally, by incorporating ABML with WCE loss, our model can acquire more semantic and discriminative features. Extensive experiments on four common data sets (CIFAR-10, large database of handwritten digits (MNIST), ImageNet-10, and NUS-WIDE) and a large-scale data set ImageNet demonstrate that our approach successfully overcomes the two difficulties above and significantly outperforms state-of-the-art hashing methods.

Large neural networks are difficult to deploy on mobile devices because of intensive computation and storage. To alleviate it, we study ternarization, a balance between efficiency and accuracy that quantizes both weights and activations into ternary values. In previous ternarized neural networks, a hard threshold Δ is introduced to determine quantization intervals. Although the selection of Δ greatly affects the training results, previous works estimate Δ via an approximation or treat it as a hyper-parameter, which is suboptimal. In this paper, we present the Soft Threshold Ternary Networks (STTN), which enables the model to automatically determine quantization intervals instead of depending on a hard threshold. Concretely, we replace the original ternary kernel with the addition of two binary kernels at training time, where ternary values are determined by the combination of two corresponding binary values. At inference time, we add up the two binary kernels to obtain a single ternary kernel. Our method dramatically outperforms current state-of-the-arts, lowering the performance gap between full-precision networks and extreme low bit networks. Experiments on ImageNet with AlexNet (Top-1 55.6%), ResNet-18 (Top-1 66.2%) achieves new state-of-the-art.

Face alignment is an important task in the field of multi-media. Together with the impressive progress of algorithms, various benchmark datasets have been released in recent years. Intuitively, it is meaningful to integrate multiple labeled datasets with different annotations to achieve higher performance on a target landmark detector. Although numerous efforts have been made in joint usage, there yet remain three shortages in recent works, e.g., additional computation, limitation of the markups scheme, and limited support for the regression method. To address the above problems, we proposed a novel Alternating Training Framework (ATF), which leverages similarity and diversity across multi-media sources for a more robust detector. Our framework mainly contains two sub-modules: Alternating Training with Decreasing Proportions (ATDP) and Mixed Branch Loss (mathcal LMB). In particular, ATDP trains multiple datasets simultaneously to take advantage of the diversity between them, while mathcal LMB utilizes similar landmark pairs to constrain different branches of corresponding datasets. Extensive experiments on various benchmarks show the effectiveness of our framework, and ATF is feasible for both heatmap-based network and direct coordinate regression. Specifically, the mean error even reaches 3.17 on the experiment on 300W leveraging WFLW, which significantly outperforms state-of-the-art methods. Both in an ordinary convolutional network (OCN) and HRNET, ATF achieves up to 9.96% relative improvement. Our source codes are made publicly available at https://github.com/starhiking/ATF.

Deep convolutional neural networks (CNNs) have shown appealing performance on various computer vision tasks in recent years. This motivates people to deploy CNNs to real-world applications. However, most of state-of-art CNNs require large memory and computational resources, which hinders the deployment on mobile devices. Recent studies show that low-bit weight representation can reduce much storage and memory demand, and also can achieve efficient network inference. To achieve this goal, we propose a novel approach named BWNH to train Binary Weight Networks via Hashing. In this paper, we first reveal the strong connection between inner-product preserving hashing and binary weight networks, and show that training binary weight networks can be intrinsically regarded as a hashing problem. Based on this perspective, we propose an alternating optimization method to learn the hash codes instead of directly learning binary weights. Extensive experiments on CIFAR10, CIFAR100 and ImageNet demonstrate that our proposed BWNH outperforms current state-of-art by a large margin.

Every bit matters in the hardware design of quantized neural networks. However, extremely-low-bit representation usually causes large accuracy drop. Thus, how to train extremely-low-bit neural networks with high accuracy is of central importance. Most existing network quantization approaches learn transformations (low-bit weights) as well as encodings (low-bit activations) simultaneously. This tight coupling makes the optimization problem difficult, and thus prevents the network from learning optimal representations. In this paper, we propose a simple yet effective Two-Step Quantization (TSQ) framework, by decomposing the network quantization problem into two steps: code learning and transformation function learning based on the learned codes. For the first step, we propose the sparse quantization method for code learning. The second step can be formulated as a non-linear least square regression problem with low-bit constraints, which can be solved efficiently in an iterative manner. Extensive experiments on CIFAR-10 and ILSVRC-12 datasets demonstrate that the proposed TSQ is effective and outperforms the state-of-the-art by a large margin. Especially, for 2-bit activation and ternary weight quantization of AlexNet, the accuracy of our TSQ drops only about 0.5 points compared with the full-precision counterpart, outperforming current state-of-the-art by more than 5 points.

Deep neural networks have evolved remarkably over the past few years and they are currently the fundamental tools of many intelligent systems. At the same time, the computational complexity and resource consumption of these networks continue to increase. This poses a significant challenge to the deployment of such networks, especially in real-time applications or on resource-limited devices. Thus, network acceleration has become a hot topic within the deep learning community. As for hardware implementation of deep neural networks, a batch of accelerators based on a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC) have been proposed in recent years. In this paper, we provide a comprehensive survey of recent advances in network acceleration, compression, and accelerator design from both algorithm and hardware points of view. Specifically, we provide a thorough analysis of each of the following topics: network pruning, low-rank approximation, network quantization, teacher–student networks, compact network design, and hardware accelerators. Finally, we introduce and discuss a few possible future directions.

We are witnessing an explosive development and widespread application of deep neural networks (DNNs) in various fields. However, DNN models, especially a convolutional neural network (CNN), usually involve massive parameters and are computationally expensive, making them extremely dependent on high-performance hardware. This prohibits their further extensions, e.g., applications on mobile devices. In this paper, we present a quantized CNN, a unified approach to accelerate and compress convolutional networks. Guided by minimizing the approximation error of individual layer's response, both fully connected and convolutional layers are carefully quantized. The inference computation can be effectively carried out on the quantized network, with much lower memory and storage consumption. Quantitative evaluation on two publicly available benchmarks demonstrates the promising performance of our approach: with comparable classification accuracy, it achieves 4 to $6 \times $ acceleration and 15 to $20\times $ compression. With our method, accurate image classification can even be directly carried out on mobile devices within 1 s.

Content-based image retrieval (CBIR) is one of the most important applications of computer vision. In recent years, there have been many important advances in the development of CBIR systems, especially Convolutional Neural Networks (CNNs) and other deep-learning techniques. On the other hand, current CNN-based CBIR systems suffer from high computational complexity of CNNs. This problem becomes more severe as mobile applications become more and more popular. The current practice is to deploy the entire CBIR systems on the server side while the client side only serves as an image provider. This architecture can increase the computational burden on the server side, which needs to process thousands of requests per second. Moreover, sending images have the potential of personal information leakage. As the need of mobile search expands, concerns about privacy are growing. In this article, we propose a fast image search framework, named DeepSearch, which makes complex image search based on CNNs feasible on mobile phones. To implement the huge computation of CNN models, we present a tensor Block Term Decomposition (BTD) approach as well as a nonlinear response reconstruction method to accelerate the CNNs involving in object detection and feature extraction. The extensive experiments on the ImageNet dataset and Alibaba Large-scale Image Search Challenge dataset show that the proposed accelerating approach BTD can significantly speed up the CNN models and further makes CNN-based image search practical on common smart phones.

K-means algorithm has been widely used in machine learning and data mining due to its simplicity and good performance. However, the standard k-means algorithm would be quite slow for clustering millions of data into thousands of or even tens of thousands of clusters. In this paper, we propose a fast k-means algorithm named multi-stage k-means (MKM) which uses a multi-stage filtering approach. The multi-stage filtering approach greatly accelerates the k-means algorithm via a coarse-to-fine search strategy. To further speed up the algorithm, hashing is introduced to accelerate the assignment step which is the most time-consuming part in k-means. Extensive experiments on several massive datasets show that the proposed algorithm can obtain up to 600X speed-up over the k-means algorithm with comparable accuracy.

Hashing methods play an important role in large scale image retrieval. Traditional hashing methods use hand-crafted features to learn hash functions, which can not capture the high level semantic information. Deep hashing algorithms use deep neural networks to learn feature representation and hash functions simultaneously. Most of these algorithms exploit supervised information to train the deep network. However, supervised information is expensive to obtain. In this paper, we propose a pseudo label based unsupervised deep discriminative hashing algorithm. First, we cluster images via K-means and the cluster labels are treated as pseudo labels. Then we train a deep hashing network with pseudo labels by minimizing the classification loss and quantization loss. Experiments on two datasets demonstrate that our unsupervised deep discriminative hashing method outperforms the state-of-art unsupervised hashing methods.

Convolutional neural networks (CNNs) have achieved impressive performance in a wide range of computer vision areas. However, the application on mobile devices remains intractable due to the high computation complexity. In this demo, we propose the Quantized CNN (Q-CNN), an efficient framework for CNN models, to fulfill efficient and accurate image classification on mobile devices. Our Q-CNN framework dramatically accelerates the computation and reduces the storage/memory consumption, so that mobile devices can independently run an ImageNet-scale CNN model. Experiments on the ILSVRC-12 dataset demonstrate 4~6x speed-up and 15~20x compression, with merely one percentage drop in the classification accuracy. Based on the Q-CNN framework, even mobile devices can accurately classify images within one second.

Recently, convolutional neural networks (CNN) have demonstrated impressive performance in various computer vision tasks. However, high performance hardware is typically indispensable for the application of CNN models due to the high computation complexity, which prohibits their further extensions. In this paper, we propose an efficient framework, namely Quantized CNN, to simultaneously speed-up the computation and reduce the storage and memory overhead of CNN models. Both filter kernels in convolutional layers and weighting matrices in fully-connected layers are quantized, aiming at minimizing the estimation error of each layer's response. Extensive experiments on the ILSVRC-12 benchmark demonstrate 4 ~ 6× speed-up and 15 ~ 20× compression with merely one percentage loss of classification accuracy. With our quantized CNN model, even mobile devices can accurately classify images within one second.

Two tasks have been put forward in the MSR-bing Grand Challenge 2015. To address the information retrieval task, we raise and integrate a series of methods with visual features obtained by convolution neural network (CNN) models. In our experiments, we discover that the ranking strategies of Hierarchical clustering and PageRank methods are mutually complementary. Another task is fine-grained classification. In contrast to basic-level recognition, fine-grained classification aims to distinguish between different breeds or species or product models, and often requires distinctions that must be conditioned on the object pose for reliable identification. Current state-of-the-art techniques rely heavily upon the use of part annotations, while the bing datasets suffer both abundance of part annotations and dirty background. In this paper, we propose a CNN-based feature representation for visual recognition only using image-level information. Our CNN model is pre-trained on a collection of clean datasets and fine-tuned on the bing datasets. Furthermore, a multi-scale training strategy is adopted by simply resizing the input images into different scales and then merging the soft-max posteriors. We then implement our method into a unified visual recognition system on Microsoft cloud service. Finally, our solution achieved top performance in both tasks of the contest