Deep Neural Networks (DNNs) are extensively utilized in various fields due to their remarkable performance. However, as DNN models increase in complexity and size, the training process incurs substantial data transfer costs between computation and storage. The slowdown of Moore’s Law further challenges the integration of additional resources on a single chip, making it difficult to improve storage capacity and reduce off-chip data transfers. To address these challenges, we propose CAMC, a multi-chiplet DNN accelerator with a heterogeneous memory computing architecture. CAMC integrates SRAM-based in-memory computing with TSV-stacked DRAM-based near-memory computing. In addition, an efficient mapping strategy was developed to optimize resource utilization and performance. The experimental results demonstrate that CAMC enhances energy efficiency by 11.84 times and reduces data transfer costs by 8.78 times compared to the baseline design.

In the data-driven era, privacy protection has become a critical concern. Differential privacy is an effective technique that incorporates random noise during data processing to ensure that alterations to individual data points do not significantly affect overall outputs. However, the additional operations required by differential privacy can result in prolonged training times and degraded model performance. This work proposes DPAcc, an FPGA-based acceleration framework for differential privacy that utilizes hardware implementation to decrease training time. The designed FPGA module efficiently executes clipping and noise addition operations, significantly reducing the overhead compared to standard training. Experimental results demonstrate that DPAcc improves training efficiency across multiple models, achieving up to 2× speedup compared to standard differential privacy training methods.

Conventional models for hyperspectral image (HSI) classification usually demand a substantial quantity of labeled training data. However, when labeled training HSIs are unavailable or have different distributions from test HSIs, many classification models tend to exhibit significant performance declines. For the cross-scene HSI classification task, unsupervised domain adaptation (UDA) technique has been widely developed. Existing discrepancy-based or adversarial-based UDA methods may fail to learn discriminative class boundaries when large class distribution shift (CDS) exists. To alleviate this limitation, we propose the model named spatially enhanced refined classifier (SERC), which includes a coarse classifier (CC) and a refined classifier (RC). The RC constructs a memory module to fuse global-spatial and spectral information simultaneously and uses the neighborhood aggregation technique to generate the refined predictions. The refined predictions are then transferred to pseudolabels to train the CC. Thereby, a mutually reinforcing relationship between the two classifiers is established. Furthermore, we propose the class distribution match (CDM) strategy to further alleviate the serious CDS problem. Notably, SERC does not require additional training parameters, which are commonly used in existing UDA methods. Despite its simplicity, SERC achieves outstanding results. Our method has been extensively evaluated on three public HSI datasets and has shown superior performance compared with state-of-the-art (SOTA) approaches. The source code can be found athttps://github.com/Wangzs0228/SERC.

Environmental perception provides the necessary information for unmanned ground vehicles to recognize and interact with surrounding objects. Velodyne light detection and ranging (LiDAR) is widely used for this purpose due to its significant advantages such as high precision and being uninfluenced by varying illuminations. However, the unstructured distribution of LiDAR point clouds always affects the performance of feature extraction and object recognition. Moreover, the numbers of parameters in most deep learning models of object recognition are very large and the training process costs lots of computation consumption. This paper proposes a broad learning system (BLS) variant with a unified space autoencoder (USAE) as a lightweight model to recognize 3D objects. When the proposed method was evaluated on the LiDAR point cloud dataset and ModelNet10 dataset, the experimental results indicated that the recognition accuracy of our USAE-BLS model was similar to that of state-of-the-art 3D object recognition models. Moreover, the USAE-BLS has a much smaller model size and shorter training time than that of the deep learning models.

Reconstructing hand-object interactions is a challenging task due to strong occlusions and complex motions. This article proposes a real-time system that uses a single depth stream to simultaneously reconstruct hand poses, object shape, and rigid/non-rigid motions. To achieve this, we first train a joint learning network to segment the hand and object in a depth image, and to predict the 3D keypoints of the hand. With most layers shared by the two tasks, computation cost is saved for the real-time performance. A hybrid dataset is constructed here to train the network with real data (to learn real-world distributions) and synthetic data (to cover variations of objects, motions, and viewpoints). Next, the depth of the two targets and the keypoints are used in a uniform optimization to reconstruct the interacting motions. Benefitting from a novel tangential contact constraint, the system not only solves the remaining ambiguities but also keeps the real-time performance. Experiments show that our system handles different hand and object shapes, various interactive motions, and moving cameras.

Software defined networking (SDN) is a promising network architecture, which decouples the control plane and data plane of a network. However, SDN opens some security challenges, such as man-in-the-middle attacks, spoofing attacks, flooding attacks and so on. In this paper, we focus on flooding attacks which consume the switch buffer and controller resource resulting in SDN framework resource overloaded. To prevent SDN framework from flooding attack, we present a defense approach called PBUF (Packet forwarding based on BUFfer sharing), which pools the idle switches to mitigate threat issues. This approach consists of buffer management and packet forwarding modules. The buffer management module gleans the statistics of incoming packets and then analyzes these statistics to estimate the buffer size by network calculus. Considering that a lot of table-miss packets will be generated and stored in buffer when the flooding attack is happening, the packet forwarding module is designed to forward these table-miss packets to idle switches to prevent the switch or controller to be overloaded. These table-miss packets will be buffered in idle switches and then sent to controller in a limited rate by generating packet_in messages. The simulation results show that PBUF is effective and only introduces a little overhead in SDN framework.