Session-based recommendation (SBR) aims to provide users with satisfactory suggestions via modeling preferences based on short-term, anonymous user-item interaction sequences. Traditional single interest learning methods struggle to align with the diverse nature of preferences. Recent advances resolved this bottleneck by learning multiple interest embeddings for each session. However, due to the pre-defining scheme of interest quantity (e.g. the number of interests), these approaches are deficient in adaptive ability towards distinctive preference patterns across different users. Moreover, these methods rely solely on the current session and ignore useful information from related ones. The short-term property of sessions would magnify the insufficient representation issue. To address these limitations, we propose a Neural Process-based Multi-interest learning framework for Session-based Recommendation, namely NP-MiSR. To be specific, our method enables adaptive multi-interest representation learning through two complementary mechanisms: 1) Neural Process-based Intra-session interest modeling: We employ Neural Processes to model the distribution of interests within a session, where the fixed interest configurations are no longer needed. 2) Cross-session context fusion: We extract interest distributions of similar sessions as contextual priors to refine the current session’s interest representation. Extensive experiments on three datasets demonstrate that our method consistently outperforms state-of-the-art SBR approaches with an average improvement of 38.8%. Moreover, the few-shot learning task reveals that NP-MiSR achieves a surprisingly favorable efficiency v.s. performance trade-off where utilizing only 10% of the training data attains 95% of the recommendation performance.

Session-based recommendation (SBR) is widely used in e-commerce and streaming services, with the task of performing real-time recommendations based on short-term anonymous user history data. Most existing SBR frameworks follow the pattern of learning a single representation for a specific session, which makes it difficult to capture potential multiple interests, thus preventing discriminative recommendations. Multi-Interest learning has emerged as an effective approach for addressing this issue on sequential data in recent years. However, the current Multi-Interest frameworks act terrible on session data because they may generate excessive interests. To address these issues, we proposed a model named Dynamic Multi-Interst Graph Neural Network (DMI-GNN) ,which introduces the Multi-Interest learning framework into SBR and refines it by proposing a multiple positional patterns (MPP) learning method and a Dynamic Multi-Interest (DMI) regularization.To be specific, the MPP learning layer ensures the model to obtain representations with different positional information for sessions.The DMI regularization, on the other hand, mitigates the influence of excessive interests. Experiments on three bench-mark datasets demonstrate that our methods achieve better performance on different metrics

Coding with hardware description languages (HDLs) such as Verilog is a time-intensive and laborious task. With the rapid advancement of large language models (LLMs), there is increasing interest in applying LLMs to assist with HDL coding. Recent efforts have demonstrated the potential of LLMs in translating natural language to traditional HDL Verilog. Chisel, a next-generation HDL based on Scala, introduces higher-level abstractions, facilitating more concise, maintainable, and scalable hardware designs. However, the potential of using LLMs for Chisel code generation remains largely unexplored. This work proposes ReChisel, an LLM-based agentic system designed to enhance the effectiveness of Chisel code generation. ReChisel incorporates a reflection mechanism to iteratively refine the quality of generated code using feedback from compilation and simulation processes, and introduces an escape mechanism to break free from non-progress loops. Experiments demonstrate that ReChisel significantly improves the success rate of Chisel code generation, achieving performance comparable to state-of-the-art LLM-based agentic systems for Verilog code generation.

Coarse-Grained Reconfigurable Arrays (CGRAs) are specialized accelerators commonly employed to boost performance in workloads with iterative structures. Existing research typically focuses on compiler or architecture optimizations aimed at improving CGRA performance, energy efficiency, flexibility, and area utilization, under the idealistic assumption that kernels can access all data from Scratchpad Memory (SPM). However, certain complex workloads–particularly in fields like graph analytics, irregular database operations, and specialized forms of high-performance computing (e.g., unstructured mesh simulations)–exhibit irregular memory access patterns that hinder CGRA utilization, sometimes dropping below 1.5%, making the CGRA memory-bound. To address this challenge, we conduct a thorough analysis of the underlying causes of performance degradation, then propose a redesigned memory subsystem and refine the memory model. With both microarchitectural and theoretical optimization, our solution can effectively manage irregular memory accesses through CGRA-specific runahead execution mechanism and cache reconfiguration techniques. Our results demonstrate that we can achieve performance comparable to the original SPM-only system while requiring only 1.27% of the storage size. The runahead execution mechanism achieves an average 3.04× speedup (up to 6.91×), with cache reconfiguration technique providing an additional 6.02% improvement, significantly enhancing CGRA performance for irregular memory access patterns.

Multibeam spaceborne photon-counting laser altimetry obtains multiple tracks simultaneously, enabling coverage of wide areas and various land-cover types. To cope with the rapid denoising of background noise in photon point cloud (PPC), this letter presents a voxel-based denoising (VoxelD) for multitrack PPC, utilizing voxel attributes and topological relationships to automatically identify signal photons within the noisy PPC. The approach involves two main steps: 1) partitioning the initial PPC into coarse resolution voxels to remove noise photons based on photon density thresholds with each voxel, reducing computational load and 2) refining the extraction process using fine resolution voxels to filter signal photon-containing voxel clusters based on three criteria: a high number of connected voxels, a topologically related voxel cluster, and enough internal photons abundance. The retained photons within the extracted voxels are considered as signal photons. The proposed method is tested with multitrack PPC from Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2)/Advanced Topographic Laser Altimeter System (ATLAS) and compared with modified density-based spatial clustering of applications with noise (DBSCAN), modified ordering points to identify the clustering structure (OPTICS), and ATL08 data product. Results demonstrate the method outperforms the compared methods, achieving precision, recall, and F score of 0.9701, 0.9530, and 0.9615, respectively.

Nearly all global datasets from satellite missions have been instrumental in advancing research on regional and global-scale underground density structures. However, satellite gravity inversion faces challenges such as instability, multiple solutions, and low resolution due to the ill-posed nature of the problem. Methods like depth-weighting and different norms improve resolution but can introduce instability and noise sensitivity, requiring further refinement for optimal results. This article proposes a novel gravity inversion method in spherical coordinates, employing dynamically re-weighting matrix to enhance inversion resolution. Initially, weights are assigned to each tesseroid based on cross-correlation coefficients, forming an initial model weighting matrix. This matrix is then iteratively optimized by minimizing two regularized objective functions that incorporate the kernel matrix and observed data, refining the weights distribution to better approximate the actual geological situation. Synthetic model studies demonstrate that this method significantly improves the resolution of inversion results, particularly in identifying gently dipping density anomalies. Application of this method to the India-Asia collision zone reveals consistent subduction plate geometry with previous studies, validating its practicality and effectiveness in imaging intricate density patterns. This approach offers a substantial advancement in gravity inversion techniques, providing clearer and more accurate geological models.

In situ detection on Mars can provide detailed information on the planet’s topography and material composition while also validating the results obtained by orbiter probes. The laser-induced breakdown spectroscopy (LIBS) has emerged as a popular technology for Mars in situ exploration due to its fast response and high accuracy in identifying elements. The analysis of LIBS data obtained by different in situ scientific payloads onboard Mars rovers can help explain scientific problems related to about Martian geological genesis and history. However, it is essential to correct the data acquired by different instruments for joint analysis and to facilitate scientific discoveries due to variances in instrument specifications and data acquisition conditions. This article presents a novel cross-instrument spectral harmonization (CISH) approach that can eliminate differences in intensity and peak positions in LIBS spectra from different instruments. In particular, a peak position consistency correction (P2C2) method is proposed to correct cross-instrument peak position inconsistency by eliminating noise or irregular bumps presented in the LIBS spectra that may be incorrectly identified as characteristic peaks. The proposed CISH approach was validated using real Mars in situ LIBS data acquired by the chemistry and camera tool (ChemCam) and Mars surface composition detector (MarSCoDe). The experimental results demonstrate increased consistency in intensity and peak positions. Specifically, the average intensity difference decreased from 3.1287 to 2.1898, and the average peak position difference decreased from 0.1540 to 0.0335 nm. Meanwhile, the accuracy of inversion after consistency correction for the same calibration target (Norite) was also improved. The average root-mean-square error (RMSE) of eight oxides decreased from 5.18 to 3.37 by using a support vector machine (SVM) and from 18.80 to 7.07 using a partial least squares-submodel (PLS-SM). The proposed approach has the potential to establish a uniform benchmark for LIBS data acquired by different instruments at different times and locations, ensuring data consistency and comparability of identified material composition results.

The belief propagation (BP) algorithm exhibits outstanding detection performance for the multiple-input multiple-output (MIMO) transmission. However, this algorithm may fail to converge because of the fully connected factor graph under the MIMO settings. To address this issue, a novel deep learning detector based on the BP algorithm (DLBP detector) is proposed, which combines the BP algorithm with the deep learning methods. The log likelihood ratios (LLR) messages are passed on the MIMO factor graph to detect the signals from the transmitters. In addition, the weights are trained via the deep learning methods and further assigned to the messages updated in the DLBP detector. Finally, simulations show that, compared with the BP detector and the damped BP detector, the DLBP detector has lower complexity and better bit error rate(BER) performance.