Foundation models, such as CNNs and ViTs, have powered the development of image representation learning. However, general guidance to model architecture design is still missing. Inspired by the connection between image representation learning and heat conduction, we model images by the heat conduction equation, where the essential idea is to conceptualize image features as temperatures and model their information interaction as the diffusion of thermal energy. Based on this idea, we find that many modern model architectures, such as residual structures, SE block, and feed-forward networks, can be interpreted from the perspective of the heat conduction equation. Therefore, we leverage the heat equation to design new and more interpretable models. As an example, we propose the Heat Conduction Layer and the Refinement Approximation Layer inspired by solving the heat conduction equation using Finite Difference Method and Fourier series, respectively. The main goal of this paper is to integrate the overall architectural design of neural networks into the theoretical framework of heat conduction. Nevertheless, our Heat Conduction Network (HcNet) still shows competitive performance, e.g., HcNet-T achieves 83.0% top-1 accuracy on ImageNet-1K while only requiring 28M parameters and 4.1G MACs. The code is publicly available at: https://github.com/ZheminZhang1/HcNet.

In real-world data, one class can contain several local clusters,e.g., birds of different poses, which makes it difficult to represent the feature distribution of each class using only a single center. Existing intra-class multimodal representation methods employ sub-centers to capture intra-class variations. However, these sub-center methods have some limitations, i.e., they ignore the relationship between sub-centers and do not ensure the diversity of sub-centers. To address these limitations, we propose a novel multi-center classifier. Different from the vanilla multi-center classifier, our proposal is established on the assumption that the deep features of the training set follow a Gaussian Mixture distribution. Specifically, we create a conditional Gaussian distribution for each class and then sample multiple sub-centers from that distribution to extend the linear classifier. This approach allows the model to capture intra-class local structures more efficiently. In addition, we propose a novel label assignment strategy, the Multi-Center Class Label, to ensure that each sub-center is effectively involved in the training. Extensive experiments on various recognition benchmarks like ImageNet, CIFAR, and Mini-ImageNet demonstrate the effectiveness of our proposal.

Vision Graph Neural Network (ViG) is the first graph neural network model capable of directly processing image data. The community primarily focuses on the model structures to improve ViG's performance but lacks attention to its graph construction method. To avoid quadratic computational complexity, ViG uses clustering algorithms (K-nearest neighbor) to construct graph structures. Nevertheless, clustering algorithms introduce biases, which limit ViG's ability to obtain global information. To address this problem, we propose RandomViG, which abandons clustering algorithms and uses a random manner to obtain relationships between nodes. Our RandomViG is sparse in computation and can approximate a complete graph, enabling ViG to gain global interaction capability. In order to obtain the local dependence, we design a local feature extraction module for RandomViG. In addition, to alleviate the over-smoothing problem, we propose a novel method called MRN (maintaining relationships among nodes). Considering that the increased feature diversity does not necessarily lead to better performance, MRN does not aim to maximize the feature diversity of the model but instead strives to maintain consistency between the feature similarity and the inherent similarity of the original image. We validate our proposal in three major computer visual tasks, including image classification, object detection, and instance segmentation. Without extra data, RandomViG-Ti achieves 79.4% ImageNet-1 K top-1 accuracy, outperforming the baseline (ViG) by 1.2%. Under the same model scale, our RandomViG performs better with fewer FLOPs compared with existing state-of-the-art models.

Positional encoding is important for vision transformer (ViT) to capture the spatial structure of the input image. General effectiveness has been proven in ViT. In our work we propose to train ViT to recognize the positional label of patches of the input image, this apparently simple task actually yields a meaningful self-supervisory task. Based on previous work on ViT positional encoding, we propose two positional labels dedicated to 2D images including absolute position and relative position. Our positional labels can be easily plugged into various current ViT variants. It can work in two ways: (a) As an auxiliary training target for vanilla ViT for better performance. (b) Combine the self-supervised ViT to provide a more powerful self-supervised signal for semantic feature learning. Experiments demonstrate that with the proposed self-supervised methods, ViT-B and Swin-B gain improvements of 1.20% (top-1 Acc) and 0.74% (top-1 Acc) on ImageNet, respectively, and 6.15% and 1.14% improvement on Mini-ImageNet. The code is publicly available at: https://github.com/zhangzhemin/PositionalLabel.

Mask image modeling (MIM) has performed excellently as a transformer-based self-supervised method via random masking and reconstruction. However, since the unmasked image patches are non-participation in the loss computation, MIM cannot effectively utilize the data and waste much computation. This drawback usually limits the learning ability of the pre-training model when pre-training on small-scale datasets. To solve this problem, we propose a novel self-supervised learning method for small-scale datasets called RotMIM. Unlike MIM, RotMIM has a different pretext task: recognizing the rotation angle that is applied to the unmasked patches. RotMIM can fully utilize data and provide a stronger self-supervised signal. Moreover, to fit RotMIM, we propose a data augmentation method called FeaMix. Our proposal ensures that the mixing area with RotMIM understands that each basic unit of semantic information in an image has the same size. This consistency guarantees clean tokenization during fine-tuning after pre-training. Our proposals outperform state-of-the-art self-supervised methods on three popular datasets, Mini-ImageNet, Caltech256, and Cifar100.

Abstract In recent years, a more robust facial feature can be learned by convolutional neural networks once introducing margins into loss functions. Those methods set a margin for each class manually to squeeze the intra‐class variations within each class equally. However, the internal feature distributions of different persons in the real world are highly unbalanced, and the distance between different identities is not uniform either. As a result, applying the same margin on all classes might not lead to higher inter‐class differences. To address this problem, this paper proposes an adaptive margin based on feature distribution to squeeze the feature interior spaces of different classes. Simultaneously, because the inter‐class margin can adequately represent the distribution of different classes in the feature space, this paper proposes a novel diversity regularization method. The regularization weights of each class are dynamically set depending on their margins. This method proposed in this paper is intuitively interpretable and can be easily applied to other classification scenarios. Experiments on current existing benchmarks have demonstrated the superiority of our method over state‐of‐the‐art competitors.

There has been an increasing interest of integrating blockchain into cyber-physical systems (CPS). The design of password hashing schemes (PHSs) is in the core of blockchain security. However, no existing PHS seems to meet both the requirements of sufficient security and small code size for blockchain-based CPSs. In this article, a novel memory-hard PHS based on the classic PBKDF2 is proposed. Evaluation results show that the proposed scheme is promising for blockchain-based CPS, as it manages to provide enhanced security in comparison to PBKDF2 with limited increase in code size.

Smart healthcare has undergone new opportunities and challenges with the arrival of the Industry 4.0 era. The intelligent imaging diagnosis system is a staple part of smart healthcare, helping doctors make clinical decisions. Nevertheless, intelligent diagnosis analysis is still confronted with the issue that it is challenging to extract effective features from the limited and high‐dimensional data, particularly in resting‐state data of amnesic mild cognitive impairment (aMCI). Furthermore, the intelligent imaging diagnosis system for aMCI is conductive to make timely predicting groups that may convert to Alzheimer’s disease (AD). To improve the system’s detection performance and reduce its data redundancy, we first develop an adaptive structure feature generation strategy (ASFGS) based on the Laplacian matrix and sparse autoencoder to obtain the structural features of brain functional network (BFN). Concurrently, we present a multiscale local feature detection strategy (MLFDS) to overcome the low utilization of local features of BFN. And finally, multiscale features, including structural features and multiscale local features, are fused by concatenation method to further improve the detection performance of aMCI system. Support vector machine based on radial basis function (RBF‐SVM) for small data learning is adopted to evaluate the effectiveness of the proposed features. Besides, we employ leave‐one‐out cross‐validation strategy to avoid the overfitting problem of classifier training process. The experiment results elucidate that the accuracy (ACC) and the area under the curve (AUC) in this work provide 86.57% and 86.36%, respectively, which outperforms the traditional methods and offers new insights for accuracy requirements of the aMCI system.

Abstract Recent methods for 2D facial landmark localization perform well on close‐to‐frontal faces, but 2D landmarks are insufficient to represent 3D structure of a facial shape. For applications that require better accuracy, such as facial motion capture and 3D shape recovery, 3DA‐2D (2D Projections of 3D Facial Annotations) is preferred. Inferring the 3D structure from a single image is an ill‐posed problem whose accuracy and robustness are not always guaranteed. This paper aims to solve accurate 2D facial landmark localization and the transformation between 2D and 3DA‐2D landmarks. One way to increase the accuracy is to input more precisely annotated facial images. The traditional cascaded regressions cannot effectively handle large or noisy training data sets. In this paper, we propose a Mini‐Batch Cascaded Regressions (MBCR) method that can iteratively train a robust model from a large data set. Benefiting from the incremental learning strategy and a small learning rate, MBCR is robust to noise in training data. We also propose a new Cross‐Dimension Annotations Conversion (CDAC) method to map facial landmarks from 2D to 3DA‐2D coordinates and vice versa. The experimental results showed that CDAC combined with MBCR outperforms the‐state‐of‐the‐art methods in 3DA‐2D facial landmark localization. Moreover, CDAC can run efficiently at up to 110 fps on a 3.4 GHz‐CPU workstation. Thus, CDAC provides a solution to transform existing 2D alignment methods into 3DA‐2D ones without slowing down the speed. Training and testing code as well as the data set can be downloaded from https://github.com/SWJTU‐3DVision/CDAC.

The prediction of traffic flow is of great significance for the prevention of accidents, the avoidance of congestion and the dispatch of command center. Considering the complexity of traffic data in reality, it is an extraordinarily challenging task to forecast accurately from historical patterns. In this paper, we propose a method based on the cascaded artificial neural network (CANN) to predict traffic flow at positions. In order to express the spatial correlation of traffic data, the actual road network distance is introduced in our model. The realworld data derived from video surveillance cameras in Xiamen is used in the experiment which is compared with five baselines. To the best of our knowledge, this is the first time that CANN is applied to forecast traffic flow. The experimental results demonstrate that the CANN method has superior performance. In addition, We also discuss the impact of some external factors such as temperature, weather and holidays on the prediction results.

In cloud computing, with full control of the underlying infrastructures, cloud providers can flexibly place user jobs on suitable physical servers and dynamically allocate computing resources to user jobs in the form of virtual machines. As a cloud provider, scheduling user jobs in a way that minimizes their completion time is important, as this can increase the utilization, productivity, or profit of a cloud. In this paper, we focus on the problem of scheduling embarrassingly parallel jobs composed of a set of independent tasks and consider energy consumption during scheduling. Our goal is to determine task placement plan and resource allocation plan for such jobs in a way that minimizes the Job Completion Time (JCT). We begin with proposing an analytical solution to the problem of optimal resource allocation with pre-determined task placement. In the following, we formulate the problem of scheduling a single job as a Non-linear Mixed Integer Programming problem and present a relaxation with an equivalent Linear Programming problem. We further propose an algorithm named TaPRA and its simplified version TaPRA-fast that solve the single job scheduling problem. Lastly, to address multiple jobs in online scheduling, we propose an online scheduler named OnTaPRA. By comparing with the start-of-the-art algorithms and schedulers via simulations, we demonstrate that TaPRA and TaPRA-fast reduce the JCT by 40-430 percent and the OnTaPRA scheduler reduces the average JCT by 60-280 percent. In addition, TaPRA-fast can be 10 times faster than TaPRA with around 5 percent performance degradation compared to TaPRA, which makes the use of TaPRA-fast very appropriate in practice.

In this paper, we propose a novel analysis method for divisible load scheduling in mesh, torus and Gaussian network, a new type of interconnection network that has the same node degree as the mesh and torus, but shorter network diameter and shorter average hop distances under equal network size. The divisible scheduling in these three networks are uniformly formulated as the Maximum Finish Time Minimization (MFTM) problem. It involves minimizing the makespan of the load distribution and processing. The MTFM problem, a relaxed MFTM problem, a linear programming problem version and a heuristic algorithm are described and solved. The first three of these problems have identical solutions. The heuristic algorithm is close in performance to the optimal solution, significantly outperforms the previously described dimensional algorithm, and has much wider application range than the previously proposed phase algorithm.

The recently proposed optical cut-through (OpCut) switch holds a great potential in achieving high energy efficiency, as it allows optical packets to cut through the switch in optical domain whenever possible, which avoids power-hungry O/E/O conversion. In the OpCut switch, to ensure in-order transmission, only optical Head-of-Line (HOL) packet of a switch flow, i.e., the stream of packets sharing the same input and output port, is allowed to cut-through the switch, and optical HOL packets are always prioritized over buffered HOL packets to achieve high cut-through ratio, which is measured by the portion of packets cutting through the switch optically. However, under such priority rule, switch flows with buffered packets are at the risk of starvation, and the OpCut switch fails to achieve 100 percent throughput for all admissible i.i.d. traffics due to the unfairness in packet scheduling. To address this two issues, in this paper we propose a delay threshold rule for packet scheduling, in which buffered packets with delays exceeding a preset delay threshold are prioritized over optical packets. In the meanwhile, the cut-through ratio is very low under heavily congested traffic due to maintaining packet order, whereas the Internet is designed to accommodate a certain degree of packet reorder, which is very common in practice due to path multiplicity. In this paper, we design a bounded-reorder packet scheduling algorithm that significantly increases the cut-through ratio of the OpCut switch while allowing a small degree of out-of-order transmission. Our extensive simulation results show that the energy efficiency of OpCut switch can be significantly improved with only a very small degree of packet reordering, which has little adverse impact on the network application performance.

As the scale of data centers and cloud computing applications increases, data center networks play a critical role in meeting the huge communication bandwidth requirement of such applications. The scalability of conventional electronic data center networks is limited by wiring complexity and reaching distance of links under fixed power budget. To overcome this problem, in this paper we propose a nonblocking hybrid switch architecture, called NEO (Nonblocking Electronic and Optical), which is able to provide nonblocking interconnections for as many as 1,000,000 servers in a data center. NEO maintains electronic interconnections for intra-pod networks, while providing interpod interconnections by optical core switches, which not only increases the scalability of the switch architecture, but also lowers the switch cost and power consumption compared to other existing optical switch architectures. We also design a packet scheduler for NEO, which adopts a credit flow control mechanism and a parallel scheduling algorithm to avoid packet loss, and provide low communication latency. Our simulation results demonstrate that NEO achieves very low average packet delay compared to other existing optical switching architectures under various traffic patterns.

The ever increasing intra-chip and inter-chip traffic load in computing systems has been pushing traditional electronic interconnects to their limit in communication bandwidth, latency, and energy consumption. In order to achieve the high bandwidth and low latency required by intra-chip and inter-chip communications and mitigate the high interconnect power dissipation, optical interconnects have been considered as a promising candidate for intra-chip and inter-chip interconnections in next generation computing systems. In addition, packet switching is an efficient switching paradigm to fully utilize the communication bandwidth. However, due to lack of random access optical memory, it is challenging to implement all-optical packet switching in optical interconnects. In this paper, we exploit bufferless routing, a special type of packet-switching, to overcome the problem of lack of random access optical buffer. More specifically, we study bufferless routing in a novel optical multichip system, called Gaussian macrochip, where embedded chips are interconnected by an optical Gaussian network. By taking advantage of the underlying Hamiltonian cycles in the Gaussian network, we design a bufferless routing algorithm for the Gaussian macrochip, which routes packets along the shortest path in the absence of deflection, and guarantees that deflected packets reach their destinations within${{N}}$hops. Our extensive simulation results demonstrate that by adopting the proposed routing algorithm, Gaussian macrochip can support much higher inter-chip communication bandwidth, has much shorter average packet delay, and is more power efficient than the previously proposed architectures for optical multichip systems.

Energy efficiency of optical packet switches (OPS) is the key to ensure the profitability of backbone network providers. However, due to lack of optical random access buffer, most optical packet switches rely on electronic buffer to resolve output contention, which requires power-hungry O/E/O conversion for all packets. The recently proposed optical cut-through (OpCut) switch holds a great potential in achieving high energy efficiency, as it allows optical packets to cut through the switch in optical domain whenever possible. The energy efficiency of OpCut switch hinges on the cut-through ratio, which is the percentage of packets that cut through the switch optically. On the other hand, it is generally desirable to maintain packet order in a switch. To achieve in-order transmission, an optical packet needs to be converted to electronic form and buffered when an earlier packet from the same flow is still in the buffer, which may lead to a low cut-through ratio. In the meanwhile, the Internet is designed to accommodate a certain degree of packet reorder, which is very common in practice due to path multiplicity. In this paper, we introduce a novel reorder metric, reorder degree, to accurately describe the extent of packet reordering, and propose a flow management scheme to bound the reorder degree of transmitted flows. We then design an efficient packet scheduling algorithm that significantly increases the cutthrough ratio of the OpCut switch while allowing a small degree of out-of-order transmission. Our extensive simulation results show that the cut-through ratio can be drastically increased with only a very small reorder degree.

With the development of multiprocessor system on chips (MPSoCs), it is expected that hundreds of computing cores will be operating on a single chip in the near future. This will require high-performance on-chip networks with very low latency to provide a communication substrate for the increasing number of cores. In this paper, we consider Gaussian on-chip networks that are of significant topological advantages over traditional mesh and torus networks in terms of diameter and average hop distance. Many applications on MPSoCs need global data movement and global control to exchange data and synchronize the execution among cores, which require all-to-all broadcast communication. In this paper, we propose an all-to-all broadcast algorithm suitable for on-chip implementation on the Gaussian network topology. The algorithm utilizes controlled message flooding based on a broadcast pattern, which can be described in a formal, generic way for each node in terms of a few simple operations and can be easily built into router hardware. Furthermore, the generic broadcast pattern also ensures a balanced traffic load in all dimensions in the network so that minimum total latency for all-to-all broadcast can be achieved. The algorithm overlaps message switching time with transmission time in a pipelined fashion to further reduce the total communication latency of all-to-all broadcast. Comparison results demonstrate the topological merits of Gaussian networks and ultralow latency of the proposed all-to-all broadcast algorithm.

Clos networks and their variations such as folded- Clos networks (fat-trees) have been widely adopted as network topologies in data center networks. Since multicast is an essential communication pattern in many cloud services, nonblocking multicast communication can ensure the high performance of such services. However, nonblocking multicast Clos networks are costly due to the large number of middle stage switches required. On the other hand, server redundancy is ubiquitous in today's data centers to provide high availability of services. In this paper, we explore server redundancy in data centers to reduce the cost of nonblocking multicast Clos data center networks (DCNs). First, we show that the sufficient nonblocking condition on the number of middle stage switches for multicast Clos DCNs can be significantly reduced, when the data center is 2-redundant, i.e., each server in the data center has exactly one redundant backup. We then investigate more general cases that the data center is k-redundant (k >; 2), and show that a higher redundancy level further reduces the cost of nonblocking multicast Clos DCNs. We also extend the result to practical data centers where servers may have different number of redundant backups depending on the availability requirement of services provided. Finally, we provide a multicast routing algorithm with linear time complexity to configure multicast connections in Clos DCNs.

All-optical packet switches (OPS) are considered as a good candidate for future ultra-fast communications as they do not require optical-electronic-optical (O/E/O) conversions. However, currently there is still no practical optical random access memory available, which makes it difficult to reduce packet loss to an acceptable level in OPS. Thus, hybrid optical/electronic switch architectures, such as the switch proposed in which we refer to as the OpCut switch in this paper, are promising alternatives due to their potential to achieve ultra-low packet loss and packet delay. Although there has been extensive work on the performance modeling of different types of electronic and all-optical switches, little work has been done for the performance modeling of hybrid switches. In this paper, we present an efficient analytical model called the aggregation model that comprehensively analyzes various performance metrics of the OpCut switch under different types of traffic. By inductively aggregating more queues in the buffer into a block, the aggregation model can achieve a polynomial complexity to the switch size. We develop the aggregation model for the OpCut switch under both Bernoulli traffic and ON-OFF Markovian traffic. The effectiveness of our model is validated by extensive simulations. The results show that the aggregation model is very accurate in all tested scenarios.