M3SPCL: Multi-stage multi-grained multi-view supervised prototypical contrastive learning.

We present a complete linear scaling method for hybrid Kohn-Sham density functional theory electronic structure calculations and demonstrate its performance. Particular attention is given to the linear scaling computation of the Kohn-Sham exchange-correlation matrix directly in sparse form within the generalized gradient approximation. The described method makes efficient use of sparse data structures at all times and scales linearly with respect to both computational time and memory usage. Benchmark calculations at the BHandHLYP/3-21G level of theory are presented for polypeptide helix molecules with up to 53 250 atoms. Threshold values for computational approximations were chosen on the basis of their impact on the occupied subspace so that the different parts of the calculations were carried out at balanced levels of accuracy. The largest calculation used 307 204 Gaussian basis functions on a single computer with 72 GB of memory. Benchmarks for three-dimensional water clusters are also included, as well as results using the 6-31G** basis set.

This paper introduces a waveform generation algorithm from three speech parameters (fundamental frequency fo, spectral envelope, and band aperiodicity). The conventional speech analysis/synthesis system based on a vocoder mainly has a waveform generator based on pitch synchronous overlap and add (PSOLA). Since it uses the fast Fourier transform (FFT) to generate the vocal cord vibration, the processing speed is proportional to the f <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">o</sub> . The algorithm also uses the spectral representation of the aperiodicity, whereas the band aperiodicity is mainly used in speech synthesis applications such as statistical parametric speech synthesis. We propose a waveform generation algorithm that reduces the computational cost and memory usage without degrading the synthesized speech. The algorithm utilizes excitation signal generation by directly using the band aperiodicity. The computational cost in a certain period is fixed because the excitation signal is filtered and processed by the overlap-add (OLA) algorithm. We used the re-synthesized speech to perform two evaluations for the processing speed and sound quality. The results showed that the sound quality of speech synthesized was almost the same by our proposed algorithm as by the conventional algorithm. The proposed algorithm can also reduce computational cost and memory usage.

Although deep compositional features have achieved an amazing performance in many application scenarios, it is not easy for them to directly tackle big data case due to their high computing time and memory usage. This paper presents a comparative evaluation for the simplification of deep compositional features by exploring the existing vector quantization and binarization techniques. Different techniques display different capabilities in similarity-preserving or memory-saving when projecting the original deep compositional features into more compact visual words or binary codes. We propose a dedicated image searching framework to evaluate all the techniques in terms of computational cost, memory usage and discrimination preserving. Extensive experiments demonstrate that it is feasible to greatly reduce computational cost and memory usage of the deep compositional features while preserving enough discriminative power. In addition, some useful conclusions are derived to guide the design of better simplifying schemes.

In this paper, a perceptually optimized three-stage rate control algorithm is proposed to reduce both the computational complexity and the memory usage over the conventional post-compression rate-distortion (PCRD) method for maintaining the visual quality. The first stage is to allocate coding bits to each code block according to the associated signal energy. The second stage approaches the final truncation point through calculating rate-distortion slopes. The third stage incorporates the evaluation of perceptible distortion into the rate-distortion control scheme of PCRD method to achieve the optimal truncation point without encoding all the image data. Simulation results show that, in both computational complexity and memory usage, the performance of the proposed rate control algorithm is superior to that of the JPEG2000 rate control algorithm while achieving the same visual quality at the target bit rate.

We apply a source‐receiver compression approach to reduce the computational time and memory usage of three‐dimensional acoustic full‐waveform inversion. By detecting and quantifying the extent of redundancy in the data, we assemble a reduced set of simultaneous sources and receivers that are weighted sums of the physical sources and receivers used in the survey. Because the number of these simultaneous sources and receivers can be significantly less than those of the physical sources and receivers, the computational time and memory usage of any inversion method can be tremendously reduced. The scheme is based on decomposing the data into their principal components using a singular value decomposition scheme and the data reduction is done through the elimination of the small eigenvalues. For the inversion method, we employ the multiplicative‐regularized Gauss‐Newton approach, in which the regularization parameter is automatically and adaptively determined by the inversion process. For the forward solver, we employ a frequency‐domain iterative solver using the perfectly matched layer absorbing boundary condition and a multigrid preconditioner. We demonstrate the advantage of our approach by using the 3D SEG/EAGE salt model.

EAAR: Efficient and Accurate Action Recognition model with enhanced spatio-temporal perception.

Lattice reduction (LR) techniques have been widely adopted to enhance the detection performance with low additional complexity in spatial multiplexed multiple-input multiple-output (MIMO) systems. However, in large MIMO systems, the required number of iterations and the corresponding latency and complexity of the existing LR algorithms increase exponentially. In this brief, we propose a low-latency LR method based on Seysen's algorithm (SA) for high-dimensional MIMO systems. The proposed algorithm introduces dual-index selection and simultaneously updates two basis pairs without the recalculation of metrics to reduce the latency and complexity. Additionally, it introduces a reduced candidate set to decrease the computational complexity and memory usage. The numerical results demonstrate that the proposed LR-aided linear detector outperforms the conventional Lenstra-Lenstra-Lovász algorithm. Furthermore, the performance of the proposed method is comparable to that of SA; it achieves reductions of 40% and 12.5% in the number of iterations and computational complexity, respectively. The proposed 65-nm CMOS implementation achieves an overall latency of 196 ns and a throughput of 25.5 M matrix/s with a clock frequency of 357.1 MHz for $8 \times 8$ MIMO systems.

Nonlinear fibre propagation and transceiver impairments limit the performance of coherent optical communication systems. Established digital signal processing techniques such as digital back-propagation and Volterra equalisation can mitigate Kerr-induced nonlinear distortions, but their computational complexity and energy consumption hinder real-time deployment. This thesis develops machine-learning and neuromorphic approaches to optical channel equalisation, focusingon performance–complexity trade-offs and practical hardware feasibility. A unified training and benchmarking framework is introduced for coherent transmission scenarios, employing Q-factor and bit-error-rate alongside implementation measures including operation counts, memory footprint, and inference latency. Within this framework, a broad class of equalisers is analysed, including multi-layer perceptrons, convolutional and recurrent neural networks, and complex-valued neural networks that process in-phase and quadrature components. Model compression is studied: pruning, quantisation, and weight clustering are jointly optimised using Bayesian optimisation to identify Pareto-efficient configurations that preserve equalisation performance while reducing computational load, memory usage, and latency. Experimental evaluations on edge-device platforms demonstrate feasibility under realistic receiver constraints. The thesis also explores hardware–software co-design. Optical phase conjugation is integrated with neural equalisers to offload part of the nonlinearity compensation to the optical domain, enabling smaller models and lower digital complexity. To address analogue noise in photonic neuromorphic computing, robustness is assessed under additive and signal-dependent noise; noise-aware training, stochastic-resonance neurons, and ensemble-based “crowd equalisation” are proposed to enhance resilience. Finally, a neuromorphic equaliser combining spiking neural networks with a streaming RWKV time-mixing module is introduced. By exploiting event-driven sparsity and constant-memory sequential processing, this architecture achieves competitive equalisation performance with reduced computational and energy requirements compared to conventional deep learning models.

In this work, we propose a framework for differentiable forward and back-projector that enables scalable, accurate, and memory-efficient gradient computation for rigid motion estimation tasks. Unlike existing approaches that rely on auto-differentiation or that are restricted to specific projector types, our method is based on a general analytical gradient formulation for forward/backprojection in the continuous domain. A key insight is that the gradients of both forward and back-projection can be expressed directly in terms of the forward and back-projection operations themselves, providing a unified gradient computation scheme across different projector types. Leveraging this analytical formulation, we develop a discretized implementation with an acceleration strategy that balances computational speed and memory usage. Simulation studies illustrate the numerical accuracy and computational efficiency of the proposed algorithm. Experiments demonstrates the effectiveness of this approach for multiple X-ray imaging tasks we conducted. In 2D/3D registration, the proposed method achieves ~8x speedup over an existing differentiable forward projector while maintaining comparable accuracy. In motion-compensated analytical reconstruction and cone-beam CT geometry calibration, the proposed method enhances image sharpness and structural fidelity on real phantom data while showing significant efficiency advantages over existing gradient-free and gradient-based solutions. The proposed differentiable projectors enable effective and efficient gradient-based solutions for X-ray imaging tasks requiring rigid motion estimation.

Multi-object tracking (MOT) in satellite videos is a challenging task due to the small size and blurry features of objects, which often lead to intermittent detection and tracking instability. Many existing object detection and tracking models often struggle with these issues, as they are not designed to effectively handle the unique characteristics of satellite videos. To address these challenges, we propose LocaLock, a joint detection and tracking framework for MOT that incorporates feature matching concepts from single object tracking (SOT) to enhance tracking stability and reduce intermittent tracking results. Specifically, LocaLock utilizes an anchor-free detection backbone for efficiency and employs a local cost volume (LCV) module to perform precise feature matching in the local area. This provides valuable object priors to the detection head, enabling the model to “lock” onto objects with greater accuracy and mitigate the instability associated with small object detection. Additionally, the local computation within the LCV module ensures low computational complexity and memory usage. Furthermore, LocaLock incorporates a novel motion flow (MoF) module to accumulate and exploit temporal information, further enhancing feature robustness and consistency across frames. Rigorous evaluations on the VISO dataset demonstrate the superior performance of LocaLock, surpassing existing methods in tracking accuracy and precision within the demanding satellite video analysis domain. Notably, LocaLock achieved state-of-the-art performance on the VISO benchmark, achieving a multi-object tracking accuracy (MOTA) of 62.6 while ensuring fast running speed.

This study proposes the word vector convolutional time-series Network (WVCTS-Net) for extracting gas feature signatures from pickled vegetables to distinguish geographical origins. First, treating electronic nose sensor responses as "word sequences" akin to NLP inputs, we introduce a word vector convolutional (WVC) block to enhance interpretability. Second, a time-domain analysis module based on gated recurrent units was designed to capture long-sequence contextual dependencies. The final WVCTS-Net integrates these components into a low-complexity gas classification framework. Compared to both generic and gas-specific models, WVCTS-Net achieves optimal performance with minimal parameters while maintaining cross-gas applicability. Structural optimizations yield a lightweight model that substantially reduces computational complexity and memory usage, demonstrating exceptional efficiency on edge devices. This work provides a robust theoretical foundation for real-time intelligent gas recognition and analysis at the terminal level.

We review space–time isogeometric analysis and in particular present and extend the unconditionally stable space–time isogeometric method introduced in Fraschini et al. (Comput Math Appl 169:205–222, 2024) for the linear acoustic wave equation to multi-patch spatial domains. This formulation incorporates a stabilization strategy that enhances numerical robustness by penalizing high-order derivatives within the variational framework. Additionally, we present a solver that exploits the Kronecker structure of the problem, eliminating the need to assemble global space–time matrices, and reduces computational overhead and memory usage. A numerical experiment is proposed, that confirms the method’s stability, accuracy, and computational efficiency, highlighting its effectiveness for large-scale wave propagation problems on complex geometries.

In recent years, methods based on deep convolutional neural networks have made great progress in the field of image super-resolution reconstruction. However, mainstream approaches generally establish many network layers, leading to high computational costs and memory usage that are unsuitable for resource-limited edge devices. To alleviate this issue, a lightweight inverted residual attention network (IRAN) is proposed to obtain better super-resolution reconstruction performance with lower parameters and computation. The main structure of the IRAN consists of a series of inverted residual attention groups (IRAGs), which are mainly composed of several inverted residual attention blocks (IRABs). IRAB effectively reduces the network parameters and computation while extracting depth features by introducing the inverted residual structure and uses the simple channel attention mechanism to learn the important channel feature information. In addition, an enhanced spatial attention mechanism is introduced at the beginning and end of IRAG to further improve the reconstruction performance of the network. The experimental results show that compared with the mainstream lightweight networks, not only the peak signal-to-noise ratio and structural similarity of image quality metrics are better but also the parameters and the computational effort of the network are lower.

SOFIE is a fast Machine Learning inference engine developed at CERN, capable of translating trained deep learning models—provided in ONNX, Keras, or PyTorch formats—into C++ code for efficient inference. The generated code has minimal dependencies, making it easily integrable into the data processing and analysis workflows of high-energy physics (HEP) experiments. This study presents a comprehensive benchmark analysis of SOFIE against leading machine learning frameworks for model evaluation, including PyTorch, TensorFlow XLA, and ONNX Runtime. We focus on evaluating their performance in HEP applications, particularly for typical models such as Graph Neural Networks for jet tagging, and Variational Autoencoders and Generative Adversarial Networks for fast simulation. Our assessment considers key factors such as computational speed, memory usage, scalability, and ease of integration with existing HEP software ecosystems. Through this comparative study, we aim to provide insights that help the HEP community select the most suitable framework for their specific needs.

Accurate time series forecasting is a cornerstone of contemporary power and energy systems, supporting essential operational and planning functions including short‐term load prediction, electricity price forecasting and so on. Transformer‐based models achieve state‐of‐the‐art results by embedding series into tokens, but their quadratic time and space complexity limits the practical applications in real‐world scenarios. In this paper, we propose to chunk the tokens before feeding them to the encoder, in which attention calculations between tokens are limited to within the same chunk. Leveraging the frequency domain characteristics of the series, we employ locality sensitive hashing (LSH) techniques to ensure that series with similar frequency patterns are assigned to the same chunk. By fixing the chunk size, our method reduces the complexity of the transformer‐based model from to . Within each chunk, distilled features are concatenated to incorporate global information. Additionally, the running score update mechanism allows our framework to adapt the chunking strategy according to the data distribution. Extensive experiments demonstrate that, across multiple real‐world datasets and models, the proposed method maintains performance comparable to the original models while substantially reducing computational complexity and memory usage.