Recently, large reasoning models have demonstrated strong mathematical and coding abilities, and deep search leverages their reasoning capabilities in challenging information retrieval tasks. Existing deep search works are generally limited to a single knowledge source, either local or the Web. However, enterprises often require private deep search systems that can leverage search tools over both local and the Web corpus. Simply training an agent equipped with multiple search tools using flat reinforcement learning (RL) is a straightforward idea, but it has problems such as low training data efficiency and poor mastery of complex tools. To address the above issue, we propose a hierarchical agentic deep search framework, HierSearch, trained with hierarchical RL. At the low level, a local deep search agent and a Web deep search agent are trained to retrieve evidence from their corresponding domains. At the high level, a planner agent coordinates low-level agents and provides the final answer. Moreover, to prevent direct answer copying and error propagation, we design a knowledge refiner that filters out hallucinations and irrelevant evidence returned by low-level agents. Experiments show that HierSearch achieves better performance compared to flat RL, and outperforms various deep search and multi-source retrieval-augmented generation baselines in six benchmarks across general, finance, and medical domains.

Schrödinger Bridge-based diffusion models have demonstrated promising performance in signal denoising. However, since ground truth signals are unavailable during the sampling process, neural networks must be employed to learn the mapping, which breaks the theoretical coupling between diffusion and sampling processes. This paper reveals a critical inconsistency between the theoretical diffusion path and the learned sampling trajectory across different frequency bands. This diffusion-sampling inconsistency directly undermines denoising effectiveness. To address this limitation, we propose the Frequency-Dependent Scheduled Schrödinger Bridge (FDSSB), which leverages power spectral density to adaptively schedule diffusion processes across frequencies. This mechanism assigns asynchronous diffusion schedules to different frequency components, correcting the diffusion schedule to better match the sampling process. As a result, FDSSB effectively mitigates the mismatch and enhances the consistency between diffusion and sampling processes. Extensive experiments demonstrate that FDSSB achieves state-of-the-art performance, with an average scale-invariant signal-to-noise ratio improvement of 7.9066 dB over competitive approaches.

Cutting-edge machine vision applications impose stringent latency and energy efficiency demands on edge devices. To address these demands, In-Sensor Computing (ISC) architectures aim to eliminate data movement overhead, while 3D RRAM technology provides the hardware foundation of high memory density and massive computing parallelism. However, existing ISC architectures rely on static resource allocation, failing to address the dynamic "shifting bottleneck" in CNNs— where early layers are compute-bound and later layers are readout-bound. To address this, we propose Re-RIS, a Reconfigurable 3D RRAM In-Sensor architecture. By dynamically switching hardware granularity between high-parallelism and high-throughput modes, Re-RIS optimizes resource utilization for varying layer characteristics. Experimental results on VGG-16 demonstrate an end-to-end latency of 0.93 ms, achieving a 75% reduction compared to static baselines, with an energy efficiency of 244.6 TOPS/W and an area efficiency of 1.85 TOPS/mm2.

Human preference alignment is critical in building powerful and reliable large language models (LLMs). However, current methods either ignore the multi-dimensionality of human preferences (e.g. helpfulness and harmlessness) or struggle with the complexity of managing multiple reward models. To address these issues, we propose Sequential Preference Optimization (SPO), a method that sequentially fine-tunes LLMs to align with multiple dimensions of human preferences. SPO avoids explicit reward modeling, directly optimizing the models to align with nuanced human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Gradient analysis is conducted to show how SPO manages to fine-tune the LLMs while maintaining alignment on previously optimized dimensions. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences and significantly outperforms the baselines.

The complexity and heterogeneity of system-on-chip (SoC) architecture keep rapidly growing and require prolonged design cycle and cost. Recent advancements in large language models (LLMs) have opened up new avenues for agile design. In this work, we present an LLM-based multi-agent assisted SoC design methodology, which utilizes LLM agents to automatically and intelligently select, integrate, and verify SoC design. We first constructed a comprehensive IP library by retrieving existing open-source IPs as foundation design resource. By leveraging collaborative LLM multi-agent, each agent is pre-configured with unique guidelines and toolsets for SoC design steps including IP selection, SoC integration, and verification. Our methodology has been applied on two SoC design cases. The generated SoCs can achieve notable up to 27.18× and 29.67× energy efficiency improvements respectively compared to SoCs generated by existing open-source platforms.

Digital elevation model (DEM) of difference (DoD) produced by unmanned aerial vehicle (UAV)-borne laser scanning (ULS) data has been one of the important methods for monitoring catchment-scale landscape change processes, while its accuracy has been limited by the lack of understanding for the spatially variable uncertainties from systematic errors and random errors included in the DoD. In this study, the DoD uncertainty derivation (DUD) method was improved by undertaking an exhaustive error analysis, estimation of residual systematic errors, and incorporating different DoD uncertainty elimination strategies, based on multitemporal ULS data acquired from a small catchment of the Chinese Loess Plateau. The adapted method was employed to estimate the soil erosion and deposition of the catchment, while the reliability of the method was verified by the volume of mass movement and the depths of gullies measured through field surveys. Results showed that mean systematic biases were 0.025, 0.008, -0.074, and 0.051 m for multitemporal point clouds, respectively. After coregistration, the corresponding systematic bias were 0.001, 0.008, -0.016, and -0.021 m, respectively. The change results showed a significant relationship with the results of mass movement and gully depths ($R^{2}~\gt 0.8$,$p~\lt 0.01$). The adapted DUD method was able to capture different erosion processes, including gully headcut retreat, gully development, mass movement, and localized deposition, while it also achieved an underestimation of the changes compared to field survey results. In the catchment, the area of human activity contributed the highest percentage of the volumetric changes, followed by the gully slope and gully bottom, and the hillslope normally contributed the lowest. Overall, the adapted DUD method provided a reliable way for estimating geomorphic changes at the catchment scale.

High-resolution topographic data acquired by remote aerial vehicles (RAVs) have facilitated the use of digital elevation model (DEM) and DEM of difference (DoD) methods for studying geomorphic processes in complex terrain. However, insufficient understanding of systematic bias and random errors for DEMs constrained the application. In this study, we comprehensively analyzed the spatial pattern and magnitude of errors (including systematic and random errors) of DEMs derived from RAV-acquired point clouds for a topographically complex area (a subcatchment of Qiaogou in the hilly and gully loess plateau (SC_QG), China). The relationships between random errors and influential factors associated with topography, point cloud density, vegetation, and interpolation algorithms were also evaluated. On this basis, an error source thresholding (EST) method was adapted through incorporating residual systematic errors and including more impacting factors in the fuzzy inference system for random error estimation. The adapted EST (AEST) method was then employed to quantify the DoD uncertainty and geomorphic changes in two small catchments with complex terrain (i.e., SC_QG and a sub-catchment of Telagou (SC_TLG) in the hilly and gully Loess Plateau, China), while the results were verified by the changes measured by terrestrial laser scanning (TLS) and erosion pins, respectively. Results showed that mean value of systematic errors of DEMs were 0.065 and 0.005 m for SC_QG and SC_TLG, while the residual errors were reduced to 0.002 and 0.001 m after co-registration, respectively. Significant statistical relationships (${p} \lt 0.01$) were found between random errors and influential factors. The erosional volume of two study sites detected by the adapted method were −252.29 and −981.07 m3 and the corresponding depositional volume were 30.57 and 1594.32 m3, respectively. The adapted method achieved a comparable pattern and magnitude of volumetric changes with TLS results, which was superior to the original EST method in SC_QG. Besides, our method showed a lower absolute error (0.034 m) compared to the original method (0.087 m) through a comparison with erosion pins measurement in the SC_TLG. Overall, the AEST method provided a reliable tool for geomorphic change detection in areas associated with complex terrain.

Large language models (LLMs) have played a pivotal role in building communicative AI, yet they encounter the challenge of efficient updates.Model editing enables the manipulation of specific knowledge memories and the behavior of language generation without retraining.However, the robustness of model editing remains an open question.This work seeks to understand the strengths and limitations of editing methods, facilitating practical applications of communicative AI.We focus on three key research questions.RQ1: Can edited LLMs behave consistently resembling communicative AI in realistic situations?RQ2: To what extent does the rephrasing of prompts lead LLMs to deviate from the edited knowledge memory?RQ3: Which knowledge features are correlated with the performance and robustness of editing?Our empirical studies uncover a substantial disparity between existing editing methods and the practical application of LLMs.On rephrased prompts that are flexible but common in realistic applications, the performance of editing experiences a significant decline.Further analysis shows that more popular knowledge is memorized better, easier to recall, and more challenging to edit effectively.

This work presents a low quantization error readout circuit with fully-charge-domain calculation for quantization and post-process of computation-in-memory (CIM)-based neural network. The contributions include: (1) A novel residual charge accumulation function is designed to achieve charge-domain summation of quantized partial sum, and reduces 38% quantization error; (2) Charge reset is introduced in the integrate & fire circuit to realize <1 LSB INL at ±7 bits and speed of 285MHz/LSB; (3) Sample & hold, current subtraction and bidirectional counter are designed to improve 3.95× energy efficiency and 2.48× area efficiency.

Non-volatile memory (NVM)-based Computation-in-memory has demonstrated a significant advantage in high-efficiency neural networks. However, the requirement of analog-to-digital converter (ADC) and post-processing circuits not only cost high energy and area but also results in high computation errors, which tradeoffs the performance boost brought by CIM. Here, we present a specific ADC and post-processing circuit of the NVM-based CIM neural network to address these issues. The main contributions include: (1) A novel residual charge accumulation function (RCA) is designed to achieve charge-domain summation of quantized partial sum and reduces 38% quantization error; (2) Charge reset is introduced in the integrate & fire circuit to realize$3.95\times $energy efficiency and$2.48\times $area efficiency. Evaluation based on the measured results of the fabricated chip shows that the VGG-11 neural network with the proposed ADC circuit can achieve a 3.28-time improvement in energy efficiency while maintaining the same network recognition rate.

Curb detection can be used to determine road boundary information, which plays a crucial role in intelligent driving. In this paper, we propose an efficient 3D curb detection network combined with the Transformer (LCDeT), which realizes efficient and stable curb extraction from the mobile laser scanning data end-to-end. Different from the most existing algorithms that project the 3D point cloud to the 2D image, like height map or density map images before processing, we directly extract the point cloud features from the 3D point cloud to avoid the loss of spatial information of the 3D point cloud. Furthermore, we introduce SpatioTemporal Window(STWin) attention operations in the Transformer module to extract continuous, smooth curb features. In the temporal dimension, we perform a cross-attention operation on the point cloud of the historical frame and the point cloud of the current frame to improve the stability and continuity of the road edge detection results between the multi-frame point clouds. In the spatial dimension, we introduce the hybrid-attention operation on the point cloud of the current frame to extract spatially related features from the axial and local positions, respectively, to improve the detection accuracy. At last, to verify the performance, we firstly test it based on the only public curb dataset of 32-line LiDAR. The proposed LCDet achieves the state-of-the-art performance, an F1 score of 97.59%, with LCDet. At the same time, it is considered that the industry lacks roadside datasets for high-resolution LiDARs. We have collected, organized and published the industry's first 128-line laser roadside dataset, NRS-Dataset, which contains 6200 frames of point clouds in Urban during daytime and nighttime. And the NRS-Dataset will be available at [github11https://github.com/STWin1/curb]. Furthermore, we test the algorithm based on this dataset. The experimental results prove that our approach achieves high accuracy and recall in complex scenarios, which further shows the higher effective and robust performance of the LCDet algorithm than previous studies.

The potential safety hazards for the power grid caused by explosion of insulators occur again and again. Thus, the detecting and monitoring of insulators on the transmission towers is vital. In the paper, a novel method was proposed to detect insulators with high resolution satellites images. First, the SuperView-1 (0.5 m) and WorldView-3 (0.3 m) scenes of Yunnan were gathered, and then gram-schmidt method was used to fusion the original images. Second, a wide deep super resolution network (WDSR) is used to enhance the images resolution by 4 times. Third, fake color output and 1% linear stretched were applied to enhance image detail. Then, an object detection neural network based on feature pyramid networks (FPN) was used to detect transmission tower. Finally, a high-resolution network (HR-Net) was used to detect insulators on the tower. For comparison, three different class weight calculation methods and online hard example mining (OHEM) training methods of HR-Net were also proposed. HR-Net-c2-ohem final achieved highest 0.8001 of F1-Score. Therefore, our proposed method is robust to detect the insulators of transmission line tower with high resolution satellites images.

In the presented article, machine learning (ML) is employed on advanced Himawari imager (AHI) to examine real-time fire and map damaged zone over Yunnan, China. The main emphasis lies in employing machine learning as an alternative to primeval thresholding, extricating, and scrutinizing thermal anomaly using infrared (IR). Firstly, Himawari brightness temperature (BT), band ratio, albedo, and BT differences are utilized to scrutinize fire break out. Then, to ensure pixels are clear and free from clouds, the Himawari cloud product is implemented. Finally, machine learning models such as random forest (RF), artificial neural network (ANN), and time-series long short-term memory (LSTM), a deep learning model, are used to precisely classify the active fire pixels and achieved accuracies of 0.96%, 0.95%, 0.92% respectively. The results evaluated using another AHI wildfire product and inter-compared with multi-source fire products.

We propose a hybrid system of combining mobile sensor nodes and fixed sensor nodes for robust and effective public management, analysis, and prediction. In the proposed system, user privacy protection (on device, edge, and cloud) is implemented via the design of multiple levels of data sensitivity protection: User activities and location data are anonymized, consented, and aggregated (via edge computing at public infrastructures) before uploading for analysis; Specially designed event tokens are distributed for information/notification flowing and matching, and sharing user sensitive data is avoided; Event tokens also link hierarchical user groups combined with blockchain based technologies. Our system facilitates matching between user private data and public data without compromising privacy.

The MapReduce framework proposed by Google to process large data sets is an efficient framework used in many areas, such as social network, scientific research, electronic business, etc. Hence, many MapReduce frameworks are proposed and implemented on different platforms. However, these MapReduce frameworks have limitations, and they cannot handle the collision problem in the map phase, and the unbalanced workload problem in the reduce phase. In this paper, an Adaptive MapReduce Framework (A-MapCG) is proposed based on the MapCG framework, to further improve the MapReduce performance on GPU platforms. Based on the experiments, we observed that for certain MapReduce applications emitting multiple Key/value (K/V) pairs for the same key, the atomic collision problem degrades the map phase performance of the MapReduce framework substantially. In addition, the workload unbalance problem wastes parallel computing resources and limits the overall reduction phase performance of the MapReduce framework on GPU platforms. A-MapCG uses segmentation table and intra-warp combination to reduce the number of collisions during the map phase. A-MapCG also adopts balanced workload assignment to improve the reduce phase performance. The proposed A-MapCG framework is evaluated on the Tesla K40 GPU hosted by Intel Core i7-4790. The case study shows that the map phase of A-MapCG achieves a speedup of 4.63 over MapCG for the test case, Word Count, with a 64MB workload. The average reduce phase speedup of A-MapCG over MapCG with parallel reductions of Word Count is 6.92. The average reduce phase speedup of A-MapCG over MapCG with serial reductions of Word Count is 4.11.

Recently, a state-of-the-art algorithm, called deep deterministic policy gradient (DDPG), has achieved good performance in many continuous control tasks in the MuJoCo simulator. To further improve the efficiency of the experience replay mechanism in DDPG and thus speeding up the training process, in this paper, a prioritized experience replay method is proposed for the DDPG algorithm, where prioritized sampling is adopted instead of uniform sampling. The proposed DDPG with prioritized experience replay is tested with an inverted pendulum task via OpenAI Gym. The experimental results show that DDPG with prioritized experience replay can reduce the training time and improve the stability of the training process, and is less sensitive to the changes of some hyperparameters such as the size of replay buffer, minibatch and the updating rate of the target network.

The bag of visual words (BOW) model has been widely applied in the field of image recognition and image classification. However, all scale‐invariant feature transform (SIFT) features are clustered to construct the visual words which result in a substantial loss of discriminative power for the visual words. The corresponding visual phrases will further render the generated BOW histogram sparse. In this study, the authors aim to improve the classification accuracy by extracting high discriminative SIFT features and feature pairs. First, high discriminative SIFT features are extracted with the within‐ and between‐class correlation coefficients. Second, the high discriminative SIFT feature pairs are selected by using minimum spanning tree and its total cost. Next, high discriminative SIFT features and feature pairs are exploited to construct the visual word dictionary and visual phrase dictionary, respectively, which are concatenated to a joint histogram with different weights. Compared with the state‐of‐the‐art BOW‐based methods, the experimental results on Caltech 101 dataset show that the proposed method has higher classification accuracy.

In most of the GPU kernel programs, the synchronization statements are inserted manually by the programmers, which is very labor intensive, and error-prone. In this paper, we propose a synchronization optimization framework to automatically insert synchronization statements into the GPU kernels at compile time, while eliminating the redundant synchronization statements. We have shown that our framework can not only insert the synchronizations correctly, but also eliminate the redundant synchronizations, which outperforms the existing compiler frameworks that introduce redundant synchronizations using the most conservative strategy. Taking the GPU kernels as the input, our framework leverages data dependence analysis to insert synchronizations. We extend CETUS, a source-to-source compiler framework, to implement our synchronization optimization framework. Experimental results show that our proposed framework achieved 100% correctness by combining extensive evaluation and manual comparison. In addition, the number of synchronization statements in GPU kernels is reduced by 32.5%, and the number of synchronization statements executed is reduced by 28.2% on average by our synchronization optimization framework compared to the original GPU kernels.

Sparse matrix vector multiplication (SpMV) is a basic building block of many scientific applications. Several GPU accelerated SpMV algorithms for the CSR format suffer from workload unbalance for irregular matrices. In this paper, we propose a new auxiliary array assisted CSR format called local segmented reduction based CSR (LSRB-CSR), which enables synchronization free preprocessing and efficient SpMV algorithm with the light weight auxiliary arrays. It is efficient for both regular matrices and irregular matrices with tiny preprocessing overhead. We compare our LSRB-CSR based SpMV algorithm with the CSR-based SpMV from cuSPARSE, the SpMV algorithm based on segmented reduction adopted by CUDPP library, and the CSR5-based SpMV algorithm for both regular and irregular sparse matrices. Compared to cuSparse, our LSRB-CSR based SpMV algorithm could improve the performance by 26% on regular matrices and up to 4750% on irregular matrices. Compared to CUDPP, our LSRB-CSR based SpMV algorithm could improve the average SpMV performance by 210% on regular matrices and 250% on irregular matrices. Our LSRB-CSR based SpMV algorithm has comparable performance as the CSR5 based SpMV algorithm for regular matrices, and achieves better performance over the CSR5 based SpMV algorithm for irregular matrices. Experimental results show that the conversion overhead from the CSR to the LSRB-CSR is only 1/10 of the overhead from the CSR to the CSR5 on average.

GP-GPUs have been used as the platform for many applications due to their powerful computation ability and massively parallel features. In this paper, we first investigate the CSR sparse matrix format, the performance of existing optimized SpMV (Sparse matrix-vector multiplication) algorithms, and analyze the memory access patterns of the SpMV algorithms. Based on the analysis of the memory access patterns, we propose a new thread scheduling technique that can take advantage of inter-warp locality and intra-warp locality simultaneously, and also can achieve memory coalescing automatically. This proposed new scheduling technique will change the memory access pattern of SpMVs significantly. The simulation results show that the performance of the SpMV using the new proposed thread scheduling technique achieves much better performance than the implementation of the SpMV optimized by other techniques.

Fingerprint registration is a critical step in fingerprint matching. Although a variety of registration alignment algorithms have been proposed, accurate fingerprint registration remains an unresolved problem. We propose a new algorithm for fingerprint registration using orientation field. This algorithm finds the correct alignment by maximization of mutual information between features extracted from orientation fields of template and input fingerprint images. Orientation field, representing the flow of ridges, is a relatively stable global feature of fingerprint images. This method uses the statistics and distribution of global feature of fingerprint images so that it is robust to image quality and local changes in images. The primary characteristic of this method is that it uses this stable global feature to align fingerprints, and that its behavior may resemble the way humans compare fingerprints. Experimental results show that the occurrence of misalignment is dramatically reduced and that registration accuracy is greatly improved at the same time, leading to enhanced matching performance.

A new method based on deformable shape models for medical image segmentation is described. Experiments for blood cell micrographs have been conducted to verify the accuracy of the shape model-based segmentation and object shape description method. The cell segmentation method does not require user input for initialization. Coherence information between cells is utilized via a globally consistent cost function. The proposed segmentation method can be used in automated analysis for images of stained blood smear and segmentation of other medical structures. A method for shape population-based retrieval is also described. Results of population-based image queries for a database of blood cell micrographs are shown.

A method for deformable shape detection and recognition is described. Deformable shape templates are used to partition the image into a globally consistent interpretation, determined in part by the minimum description length principle. Statistical shape models enforce the prior probabilities on global, parametric deformations for each object class. Once trained, the system autonomously segments deformed shapes from the background, while not merging them with adjacent objects or shadows. The formulation can be used to group image regions obtained via any region segmentation algorithm, e.g., texture, color, or motion. The recovered shape models can be used directly in object recognition. Experiments with color imagery are reported.

A method for deformable shape detection and recognition is described. Deformable shape templates are used to partition the image into a globally consistent interpretation, determined in part by the minimum description length principle. Statistical shape models enforce the prior probabilities on global, parametric deformations for each object class. Once trained, the system autonomously segments deformed shapes from the background, while not merging them with adjacent objects or shadows. The formulation can be used to group image regions based on any image homogeneity predicate; e.g., texture, color or motion. The recovered shape models can be used directly in object recognition. Experiments with color imagery are reported.