The performance improvement by CMOS circuit technology is reaching its limits. Many researchers have been studying computing technologies that use emerging devices to challenge such critical issues. Nanophotonic technology is a promising candidate for tackling the issue due to its ultra-low latency, high bandwidth, and low power characteristics. Although previous research develops hardware accelerators by exploiting nanophotonic circuits for AI inference applications, there has never been considered for the acceleration of training that requires complex Floating-Point (FP) operations. In particular, the design balance between optical and electrical circuits has a critical impact on the latency, energy, and accuracy of the arithmetic system, and thus requires careful consideration of the optimal design. In this study, we design three types of Opto-Electrical Floating-point Multipliers (OEFMs): accuracy-oriented (Ao-OEFM), latency-oriented (Lo-OEFM), and energy-oriented (Eo-OEFM). Based on our evaluation, we confirm that Ao-OEFM has high noise resistance, and Lo-OEFM and Eo-OEFM still have sufficient calculation accuracy. Compared to conventional electrical circuits, Lo-OEFM achieves an 87% reduction in latency, and Eo-OEFM reduces energy consumption by 42%.

Annealing machines use an Ising model to represent combinatorial optimization problems (COPs) and minimize the energy of the model with spin-flip sequences. Pseudo temperature is a key hyperparameter to control the search performance of annealing machines. In general, the temperature is statically scheduled such that it is gradually decreased from a sufficiently high to a sufficiently low values. However, the search process during high and low temperatures in solving constrained COPs does not improve the solution quality as expected, which requires repeated preliminary annealing for pre-tuning. This paper proposes a flip-count-based dynamic temperature control (FDTC) method to make the preliminary annealing unnecessary. FDTC checks whether the current temperature is effective by evaluating the average number of flipped spins in a series of steps. The simulation results for traveling salesman problems and quadratic assignment problems demonstrate that FDTC can obtain comparable or higher solution quality than the static temperature scheduling pre-tuned for every COP.

Federated Learning (FL) facilitates deep learning model training across distributed networks while ensuring data privacy. When deployed on edge devices, network pruning becomes essential due to the constraints of computational resources. However, traditional FL pruning methods face bias issues arising from the varied distribution of local data, which poses a significant challenge. To address this, we propose DDPruneFL, an innovative FL pruning framework that utilizes Discriminative Data (DD). Specifically, we utilize minimally pre-trained local models, allowing each client to extract semantic concepts as DD, which then inform an iterative pruning process. As a result, DDPruneFL significantly outperforms existing methods on four benchmark datasets, adeptly handling both IID and non-IID distributions and Client Selection scenarios. This model achieves state-of-the-art (SOTA) performance in this field. Moreover, our studies comprehensively validate the effectiveness of DD. Furthermore, a detailed computational complexity analysis focused on Floating-point Operations (FLOPs) is also conducted. The FLOPs analysis reveals that DDPruneFL significantly improves performance during inference while only marginally increasing training costs. Additionally, it exhibits a cost advantage in inference when compared to other pruning FL methods of the same type, further emphasizing its cost-effectiveness and practicality.

Big data processing is a set of techniques or programming models, which can be deployed on both the cloud servers or edge nodes, to access large-scale data and extract useful information for supporting and providing decisions. Meanwhile, several typical domains of human activity in smart society, such as social networks, medical diagnosis, recommendation systems, transportation, and Internet of Things (IoT), often manage a vast collection of entities with various relationships, which can be naturally represented by the graph data structure. As one of the convincing solutions to carry out analytics for big data, graph processing is especially applicable for these application domains. However, either the intra-device or the inter-device data processing in the edge-cloud architecture is truly prone to be attacked by the malicious Trojans covertly embedded in the counterfeit processing systems developed by some third-party vendors in numerous practical scenarios, leading to identity theft, misjudgment, privacy disclosure, and so on. In this paper, for the first time to our knowledge, we specially build a novel attack model for ubiquitous graph processing in detail, which also has easy scalability for other applications in big data processing, and discuss some common existing mitigations accordingly. Multiple activation mechanisms of Trojans designed in our attack model effectively make the attacks imperceptible to users. Evaluations indicate that the proposed Trojans are highly competitive in stealthiness with trivial extra latency.

Mainstream data augmentation techniques involving image-level manipulation operations (e.g., CutMix) compromise the integrity of extracted features, which impedes the application of data augmentation for pixel-level image processing tasks. Moreover, the unexplored potential of test-time augmentation within the HDR domain remains to be validated. In this paper, a full stage data augmentation method called HDR-VDA for HDR video reconstruction is proposed, especially for synthetic video based training datasets. In the training stage, the local area-based mixed data augmentation (LMDA) provides samples encompassing diverse exposure and color patterns, thus the trained model gains improved capabilities in effectively processing poorly-exposure regions, with particular emphasis on areas with rich color details. A motion and ill-exposure guided sample rank and adjustment strategy (MISRA) is utilized to augment specific training samples and compensate extra information. In the testing stage, an HDR-targeted test-time augmentation method (HTTA) is designed for reconstructed HDR frames. After restoring the shape of the test-time augmented HDR output to be consistent with the original inference output, an ill-exposure outlier removal based average ensemble method is used to blend all augmented inference outputs to generate reliable and stable reconstruction results. Experiments demonstrate that HDR-VDA achieves a better PSNR-T score of 38.91 dB, compared with conventional works under the same conditions.

We show that a limited dataset with fixed camera scene detection and a specific frame extraction method contribute to efficient training of the CAIN model for frame interpolation. Specifically, we use the detection of the scene of fixed cameras and the training dataset limited by several conditions, i.e., the condition of triplet consecutive frame extraction.

The study of the spread of pandemics, including COVID-19, is an emerging concern to promote self-care management through social distancing, using state-of-the-art tools and technologies. Existing technologies provide many opportunities to acquire and process large volumes of data to monitor user activities from various perspectives. However, determining disease hotspots remains an open challenge considering user activities and interactions; providing related recommendations to susceptible individuals requires attention. In this article, we propose an approach to determine disease hotspots by modeling users’ activities from both cyber- and real-world spaces. Our approach uniquely connects cyber- and physical-world activities to predict hazardous regions. The availability of such an exciting data set is a non-trivial task; therefore, we produce the data set with much hard work and release it to the broader research community to facilitate further research findings. Once the data set is generated, we model it as a directed multi-attributed and weighted graph to apply classical machine learning and graph neural networks for prediction purposes. Our contribution includes mapping user events from cyber- and physical-world aspects, knowledge extraction, dataset generation, and reasoning at various levels. Within our unique graph model, numerous elements of lifestyle parameters are measured and processed to gain deep insight into a person’s status. As a result, the proposed solution enables the authorities of any pandemic, such as COVID-19, to monitor and take measurable actions to prevent the spread of such a disease and keep the public informed of the probability of catching it.

The bicube is derived from the hypercube, and it provides a topology for interconnection networks of parallel systems. The bicube can interconnect the same number of nodes with the same degree as the hypercube while its diameter is almost half of that of the hypercube. In addition, the bicube preserves the property of node symmetry. Hence, the bicube attracts much attention. In this paper, we focus on the bicube with faulty nodes and propose three fault-tolerant routing methods to find a fault-free path between any pair of non-faulty nodes in it.

Visible-infrared person re-identification (VI-ReID) aims to achieve cross-modality matching between the visible and infrared modalities, thus enabling usage in all-day monitoring scenarios. Existing VI-ReID methods have indeed achieved promising performance by considering the global information for identity-related discriminative learning. However, they often overlook the importance of local information, which can contribute significantly to learning identity-specific discriminative cues. Moreover, the substantial modality gap typically poses challenges during the model training process. In response to the aforementioned issues, we propose a VI-ReID method called partial enhancement and channel aggregation (PECA) and make efforts in the following three aspects. Firstly, to capture local information, we introduce the global-local similarity learning (GSL) module, which compels the encoder to focus on fine-grained details by increasing the similarity between global and local features within various feature spaces. Secondly, to address the modality gap, we propose an inter-modality channel aggregation learning (ICAL) approach, which progressively guides the learning of modality-invariant features. ICAL not only progressively alleviates modality gap but also augments the training data. Additionally, we introduce a novel instance-modality contrastive loss, which facilitates the learning of modality-invariant and identity-related features at both the instance and modality levels. Extensive experiments on the SYSU-MM01 and RegDB datasets have shown that PECA outperforms state-of-the-art methods.

This paper develops certified defenses for deep neural network (DNN) based content-based image retrieval (CBIR) against adversarial examples (AXs). Previous works put their effort into certified defense for classification to improve certified robustness, which guarantees that no AX to cause misclassification exists around the sample. Such certified defense, however, could not be applied to CBIR directly because the goals of adversarial attack against classification and CBIR are completely different. To develop the certified defense for CBIR, we first define the new certified robustness of CBIR, which guarantees that no AX that changes the ranking results of CBIR exists around the input images. Then, we propose computationally tractable verification algorithms that verify whether a given feature extraction DNN satisfies the certified robustness of CBIR at given input images. Our proposed verification algorithms are achieved by evaluating the upper and lower bounds of distances between feature representations of perturbed and non-perturbed images in deterministic and probabilistic manners. Finally, we propose robust training methods to obtain feature extraction DNNs that increase the number of inputs that satisfy the certified robustness of CBIR by tightening the upper and lower bounds. We experimentally show that our proposed certified defenses can guarantee robustness deterministically and probabilistically on various datasets.

This letter presents a solution for large classroom interactions using cloud computing and mobile devices. A lecturer can collect student photos or texts and give real-time feedback. Students confirmed in anonymous surveys that this solution enabled them to actively participate in classes and enhanced their learning even in large classrooms.