HyperMetric: Efficient Hyperdimensional Computing With Metric Learning for Robust Edge Intelligence

Hyperdimensional Computing (HDC) offers a robust and energy-efficient paradigm for edge intelligence; however, current hardware accelerators are often proprietary, tailored to the target learning task and tightly coupled to specific CPU microarchitectures, limiting portability and adoption. To address this, and democratize the deployment of HDC hardware, we present a general-purpose, plug-and-play accelerator IP that implements the Binary Spatter Code framework as a standalone, host-agnostic module. The design is compliant with the AMBA AXI4 standard and provides an AXI4-Lite control plane and DMA-driven AXI4-Stream datapaths coupled to a banked scratchpad memory. The architecture supports synthesis-time scalability, enabling high-throughput transfers independently of the host processor, while employing microarchitectural optimizations to minimize silicon area. A multi-layer C++ software (GitHub repository commit 3ae3b46) stack running in Linux userspace provides a unified programming model, abstracting low-level hardware interactions and enabling the composition of complex HDC pipelines. Implemented on a Xilinx Zynq XC7Z020 SoC, the accelerator achieves substantial gains over an ARM Cortex-A9 baseline, with primitive-level speedups of up to 431×. On end-to-end classification benchmarks, the system delivers average speedups of 68.45× for training and 93.34× for inference. The complete RTL and software stack are released as open-source hardware to support reproducible research and rapid adoption on heterogeneous SoCs.

Hyperdimensional computing (HD) is an emerging paradigm for machine learning based on the evidence that the brain computes on high-dimensional, distributed, representations of data. The main operation of HD is encoding, which transfers the input data to hyperspace by mapping each input feature to a hypervector, followed by a bundling procedure that adds up the hypervectors to realize the encoding hypervector. The operations of HD are simple and highly parallelizable, but the large number of operations hampers the efficiency of HD in embedded domain. In this paper, we propose SHEARer, an algorithmhardware co-optimization to improve the performance and energy consumption of HD computing. We gain insight from a prudent scheme of approximating the hypervectors that, thanks to error resiliency of HD, has minimal impact on accuracy while provides high prospect for hardware optimization. Unlike previous works that generate the encoding hypervectors in full precision and then and then perform ex-post quantization, we compute the encoding hypervectors in an approximate manner that saves resources yet affords high accuracy. We also propose a novel FPGA architecture that achieves striking performance through massive parallelism with low power consumption. Moreover, we develop a software framework that enables training HD models by emulating the proposed approximate encodings. The FPGA implementation of SHEARer achieves an average throughput boost of 104,904× (15.7×) and energy savings of up to 56,044× (301×) compared to state-of-the-art encoding methods implemented on Raspberry Pi 3 (GeForce GTX 1080 Ti) using practical machine learning datasets.

Errors in data are a key challenge in modern data management and processing systems. Monitoring and mitigating risks associated with errors in data transformations and downstream applications, such as Machine Learning (ML) model training, requires a profound understanding of error generation and impact of errors on data pipelines. Unfortunately, scientific progress in the field is facing two main challenges: For one, research on data errors often does not adhere to the FAIR (Findable, Accessible, Interoperable, and Reusable) principles, which impedes reproducibility and comparisons. Second, existing data error models are oversimplified and fail to capture the complex statistical dependencies underlying the types and distributions of errors observed in real-world data. Building on prior work in the database management systems and statistics literature, we extend the theory on missing values to encompass a broader range of errors in tables and provide an overview of relevant error types. Combining error sampling mechanisms often observed in real data with a comprehensive categorization of errors, we introduce a latent factor model for tabular data errors that is simple to implement and can effectively model realistic error dependencies. Error sampling is decoupled from error types, which allows for simple extensions with more error types or sampling mechanisms. Using established benchmarks, we evaluate our model in two application scenarios, data cleaning and tabular ML tasks. In a comprehensive suite of experiments we demonstrate the impact of realistic error models on data cleaning benchmarks. Our results also show that a simple generative error model captures a wide range of error mechanisms and offers a convenient formalization of data perturbations to improve the generalizability, robustness and reproducibility of data cleaning research.

Using low-rank adaptation to fine-tune pre-trained neural network models has attracted widespread attention due to its advantages of low resource requirements, high precision, and no additional inference delay. However, most existing methods are designed for large language models based on the Transformer structure and lack adaptation to convolutional neural networks (CNN) widely used on Internet of Things (IoT) devices. In addition, IoT devices are usually deployed outdoors and collect large amounts of data that are affected by environmental conditions. Providing a highly robust CNN model is a prerequisite for providing high-quality services. To this end, this paper proposes a highly robust seed-filter adaptation method (SeFA) for pre-trained CNNs. SeFA introduces an adaptation branch with the same structure as the backbone network. In the adaptation branch, some filters are first designated seed filters and grouped. Then, additional filters are generated from the grouped seed filters and nonlinear transformation functions (NLFs) with different hyperparameters. The seed filters' parameters are updated during model training, and the NLF hyperparameters are randomly initialized and frozen. Both grouping seed filters and configuring NLFs with non-learnable hyperparameters can improve the model's robustness. This is because grouping seed filters can generate diverse filters on demand without increasing the model's complexity, and the NLFs' rules can regularize the model. It is worth mentioning that the number of fine-tuning parameters of the pre-trained model that can adapt to downstream tasks can be flexibly controlled by specifying the number of seed filters. The key idea of this paper is to propose SeFA with flexible, controllable, learnable parameters, high robustness, and adaptability to pretrained CNN models, thereby facilitating fine-tuning on resource-constrained IoT devices and providing highly robust visual services. Experimental results on the CIFAR-10, CIFAR-10-C, CIFAR-100, CIFAR-100-C, and Icons-50 datasets demonstrate that the proposed SeFA outperforms other state-of-the-art methods. Specifically, on the ResNet-152 model and the CIFAR-10-C dataset, the accuracy of our SeFA is about 7% higher than that of the full fine-tuning method. In addition, we verify the efficiency of SeFA-based models for fine-tuning and inference on the device.

Traditional news false information detection can no longer adapt to the current information detection evasion mode. This article addressed the language changes and diversity in identifying news false information, and improved the accuracy of false information detection. Firstly, the system goal was defined to improve the accuracy of false information detection. Then, news data from FakeNewNet, BuzzFeedNews, and PoliFact platforms were collected, and the data was subjected to data cleaning, segmentation, removal of stop words, One-Hot encoding, TF-IDF (Term Frequency-Inverse Document Frequency) feature extraction, and word embedding preprocessing. After establishing a Hybrid Neural Network (HNN) model, model training, evaluation, and optimization work were carried out. In the experimental stage, 5 new datasets were added and 5 other detection algorithm models were applied to compare the detection accuracy with the proposed model. Robustness experiments were conducted to verify the robustness of the model. The experimental results showed that the accuracy of the model in this article on eight news information datasets ranged from 0.9972 to 0.9998, with a mean of 0.9987. The average accuracy of the other five algorithms was 0.8010, 0.7738, 0.7394, 0.8676, and 0.7689, respectively. The detection accuracy of the algorithm in this article was much higher than the other five algorithms, and it had high robustness. Intelligent information processing and hybrid networks have brought a more comprehensive and accurate solution to the detection of false information in news dissemination, successfully solving the problem of identifying the changing styles of false information, improving the accuracy of fake news detection, and providing a very good idea for fake news detection.

Recently, multi-bit slicing has emerged as a promising technique to improve the energy efficiency of charge-domain Compute-In-Memory (CIM) accelerators by reducing the number of Analog-to-Digital (A/D) conversions. However, multi-bit slicing requires shift-and-add operations to reconstruct outputs, which exponentially amplify errors and cause significant accuracy degradation. Existing works mainly rely on hardware-aware retraining or noise-suppression techniques, incurring considerable design or power overhead. Thus, multi-bit CIM designs often face the dilemma of trading off energy efficiency for error resilience. In this paper, we propose Cambricon-CIM, a charge-domain multi-bit CIM accelerator that achieves both high energy efficiency and strong error resilience, without requiring retraining. The core insight is that the error amplification is proportional to digit weights; and by redefining these digit weights with smaller non-binary coding bases, it is possible to reduce the total error amplification. Leveraging this principle, CambriconCIM dynamically selects the minimal coding bases for every analog dot-product. With novel circuit and architectural support, Cambricon-CIM enables fast, low-overhead reconfiguration of coding bases at runtime. Experimental results show that Cambricon-CIM achieves 2.27× energy efficiency and 3.06× performance over RAELLA, a state-of-the-art error-resilient multi-bit slicing CIM architecture.

Fine-tuning of large language models is often demanding in terms of computational resources and memory. Consequently, there is a need to explore new methods that can effectively fine-tune these models without compromising performance. In this paper, we compared traditional and parameter-efficient fine-tuning of large language model XLM-RoBERTa, across both high-resourced languages like English and low-resourced languages such as Serbian. The model is fine-tuned for sentiment analysis when it is conducted using text classification. The experiments have shown that parameter-efficient fine-tuning can produce results that are on par with traditional fine-tuning, both for the English and the Serbian language. Furthermore, parameter efficient fine-tuning approach significantly reduces the computational time and memory requirements associated with model training, thereby offering a promising avenue for resource-efficient large language models fine-tuning. In experiments with the English language, we achieved accuracy using the PEFT approach that was only 3% lower than the accuracy obtained through the traditional approach, while reducing training time by 35% and memory requirements by 42%. In experiments involving the Serbian language, we achieved accuracy using the PEFT approach that was only 1% lower than the accuracy obtained through the traditional approach, while reducing training time by 42% and memory requirements by 69%.

For the protection of trained deep neural network(DNN) models, embedding watermarks into the weights of the DNN model have been considered. However, the amount of change in the weights is large in the conventional methods, and it is reported that the existence of hidden watermark can be detected from the analysis of weight variance. This helps attackers to modify the watermark by effectively adding noise to the weight. In this paper, we focus on the fully-connected layers of fine-tuning models and apply a quantization-based watermarking method to the weights sampled from the layers. The advantage of the proposed method is that the change caused by watermark embedding is much smaller and the distortion converges gradually without using any loss function. The validity of the proposed method was evaluated by varying the conditions during the training of DNN model. The results shows the impact of training for DNN model, effectiveness of the embedding method, and high robustness against pruning attacks.

Training of large-scale deep learning models necessitates parallelizing the model and data across numerous devices, and the choice of parallelism strategy substantially depends on the training workloads such as memory consumption, computation cost, and communication cost. Current approaches generally assume uniform training workloads across samples in a given task. Thus, existing systems are designed to adopt a static parallelism strategy throughout one training process. Nevertheless, when training models with sequence inputs, this assumption fails due to the sequence length variation across samples. Consequently, training with a static parallelism strategy would result in sub-optimal performance.

Forest fire susceptibility prediction is essential for effective management, yet considerable uncertainty persists under future climate change, especially in climate-sensitive plateau regions. This study integrated MODIS fire data with climatic, topographic, vegetation, and anthropogenic variables to construct an Extreme Gradient Boosting (XGBoost) model for the Yunnan Plateau, a region highly prone to forest fires. Compared with Support Vector Machine and Random Forest models, XGBoost showed superior ability to capture nonlinear relationships and delivered the best performance, achieving an AUC of 0.907 and an overall accuracy of 0.831. The trained model was applied to climate projections under SSP1-2.6, SSP2-4.5, and SSP5-8.5 to assess future fire susceptibility. Results indicated that high-susceptibility periods primarily occur in winter and spring, driven by minimum temperature, average temperature, and precipitation. High-susceptibility areas are concentrated in dry-hot valleys and mountain basins with elevated temperatures and dense human activity. Under future climate scenarios, both the probability and spatial extent of forest fires are projected to increase, with a marked expansion after 2050, especially under SSP5-8.5. Although the XGBoost model demonstrates strong generalizability for plateau regions, uncertainties remain due to static vegetation, coarse anthropogenic data, and differences among climate models.

This paper presents an AI-driven multilingual hand sign recognition system that integrates computer vision, deep learning, and natural language processing to enable real-time gesture interpretation and cross-lingual communication. The proposed system addresses the limitations of conventional sign language interpretation methods by providing an automated, scalable, and language-independent framework for translating hand gestures into both textual and auditory outputs. The system utilizes a vision-based processing pipeline in which input video streams are analyzed using MediaPipe for efficient hand detection, localization, and three-dimensional landmark extraction. The extracted spatial features are fed into a Convolutional Neural Network (CNN) optimized for multi-class gesture classification. To enhance model generalization and reduce overfitting, the network is trained on a structured dataset of annotated hand gestures using data augmentation techniques such as rotation, scaling, and normalization. Model training is performed using categorical cross-entropy loss and adaptive optimization algorithms to ensure stable and efficient convergence. Following classification, the predicted gesture labels are mapped to semantic text representations, which are then processed through a neural machine translation module to generate multilingual outputs. A text-to-speech (TTS) synthesis component further converts the translated text into speech, enabling multimodal interaction. The system is designed to operate under real-time constraints with low computational latency, making it suitable for practical deployment in assistive environments. Experimental results demonstrate that the proposed framework achieves high recognition accuracy and robustness across varying environmental conditions. The combination of lightweight landmark-based feature extraction and deep learning-based classification significantly improves computational efficiency compared to traditional image-based methods. Overall, the proposed system contributes to the advancement of assistive technologies by providing a comprehensive, real-time, and scalable solution for inclusive communication.

Power dissipation and tolerance to process variations pose conflicting design requirements. Scaling of voltage is associated with larger variations, while Vdd upscaling or transistor up-sizing for process tolerance can be detrimental for power dissipation. However, for certain signal processing systems such as those used in color image processing, we noted that effective trade-offs can be achieved between Vdd scaling, process tolerance and "output quality". In this paper we demonstrate how these tradeoffs can be effectively utilized in the development of novel low-power variation tolerant architectures for color interpolation. The proposed architecture supports a graceful degradation in the PSNR (Peak Signal to Noise Ratio) under aggressive voltage scaling as well as extreme process variations in sub-70nm technologies. This is achieved by exploiting the fact that some computations are more important and contribute more to the PSNR improvement compared to the others. The computations are mapped to the hardware in such a way that only the less important computations are affected by Vdd-scaling and process variations. Simulation results show that even at a scaled voltage of 60% of nominal Vdd value, our design provides reasonable image PSNR with 69% power savings

Promoted by integrating sensing, communication, and computing for edge intelligence in 6 G networks, the Digital Twin (DT), as a typical 6G application, has been widely employed for minimizing the disruption of Industrial Internet-of-Things (IIoT) operations and facilitating the cost-effective execution through conducting equivalent model training. Nevertheless, in complex and changeable industrial scenarios, the existing approaches of DT model construction and update within the Digital Twin Edge Network (DITEN), do not fully consider the heterogeneity in computing and communication capabilities, resulting in unbalanced model training cost and loss, which also cannot respond quickly to the environmental changes and ensure the data consistency of virtual-real space. To realize the collaborative and efficient utilization of heterogeneous computing and communication resources, and strengthen the balance of training cost and loss in the whole life cycle of DT model, an approach of using semi-asynchronous federated learning (FL) for DT construction and update, named Saei-DT, is proposed, which includes the three significant operations, e.g., terminal selection for model training, self-adaption of local training iterations, and semi-asynchronous global aggregation. Through conducting extensive experiments on the dataset of Industrial-Equipment contained in Kaggle, the experimental results demonstrate the efficiency of Saei-DT in both training accuracy and loss.

In this paper, a new rate---distortion (R---D) optimized error-resilient scheme is proposed to improve error resilience for multi-view video transmission over lossy networks. Based on the study on the characteristics of multi-view video coding and the propagating behavior of channel errors, a recursive model to estimate the end-to-end distortion is first developed, in which the channel-induced distortion takes into consideration both motion and disparity compensation prediction. Then, based on the estimated distortion and the derived total bit rate, R---D optimized motion and disparity estimation are employed to find the reliable motion vector and disparity vector to achieve the optimum trade-off between prediction efficiency and error robustness. Finally, R---D optimization is applied again to perform coding mode decision, which explicitly considers intra, inter and inter-view prediction modes. Extensive experimental results demonstrate significant performance gains can be achieved for multi-view video communications against transmission errors.

This work is concerned with increasing the error resilience of two description scalar quantizers by applying a permutation to the index of each description, for an m-diagonal initial index assignment. First we address the existence of permutations achieving a minimum Hamming distance of at least 2. Such permutations allow the central decoder to detect any 1 bit error pattern. We establish the connection with the hypercube antibandwidth problem, connection which allows us to determine the highest value of m for which such a permutation exists and to show its construction. Further, we highlight the relation between the error robustness at the side decoders and the bandwidth of the hypercube labeling associated to the permutation. To ensure error resilience at both the central and side decoders we are interested in labelings with the lowest bandwidth given that the antibandwidth is larger or equal to m. We make some progress toward the solution of this problem by constructing a class of hypercube labelings trading the increase in antibandwidth for the decrease in bandwidth.