Large language models (LLMs) have recently been adopted for recommendation by framing user preference modeling as a language generation problem. However, existing latent reasoning approaches typically represent user intent with a single latent vector, which struggles to capture the inherently multi-faceted nature of user preferences. We propose Factorized Latent Reasoning (FLR), a novel framework for LLM-based sequential recommendation that decomposes latent reasoning into multiple disentangled preference factors. FLR introduces a lightweight multi-factor attention module that iteratively refines a latent thought representation, where each factor attends to distinct aspects of the user's interaction history. To encourage diversity and specialization, we design orthogonality, attention diversity, and sparsity regularization objectives, and dynamically aggregate factor contributions for the final prediction. We further integrate FLR with an efficient reinforcement learning strategy based on group-relative policy optimization, enabling stable alignment directly in the latent reasoning space. Experiments on multiple benchmarks show that FLR consistently outperforms strong baselines while improving robustness and interpretability. Our data and code are available at https://github.com/ToAdventure/FLR.

Recently, the field of hyperspectral image (HSI) classification has witnessed advancements with the emergence of deep learning models. Promising approaches, such as self-supervised strategies and domain adaptation, have effectively tackled the overfitting challenges posed by limited labeled samples in HSI classification. To extract comprehensive semantic information from different types of auxiliary tasks, which view the problem from multiple perspectives, and efficiently integrate multiple tasks into a single network, this paper proposes a hybrid multi-task learning framework (HyMuT) by sharing representations across multiple tasks. Based on the similarity between the data and target classification task, we construct three auxiliary tasks that are similar, related and weakly correlated to the target task, while three corresponding multi-task learning methods are integrated. The framework establishes a backbone network with a hard parameter sharing mechanism, which handles the main task and a similar spatial mask classification task. Subsequently, a hierarchical transfer multi-task learning approach is introduced to transfer the knowledge of a spatial-spectral joint mask reconstruction task from the autoencoder to the backbone network. Furthermore, a new source domain HSI dataset is introduced as an auxiliary task weakly correlated. To solve the source domain classification task and assist the hard parameter sharing mechanism, a dual adversarial classifier based on adversarial learning is employed. This classifier effectively extracts domain and task invariance. Extensive experiments are conducted on four benchmark HSI datasets to evaluate the performance. The results demonstrate that HyMuT outperforms state-of-the-art methods. This code will be available from the website: https://github.com/HaoLiu-XDU/HyMuT.

Change detection (CD) methods have been crucial in exploring geo-environmental science. With the advancement of remote sensing (RS) technology, multimodal images acquired from different platforms and sensors are widely used for CD tasks. As an emerging task, multimodal CD (MCD) aims to achieve more comprehensive and precise detection of land cover changes through complementary information in multimodal images. However, there are significant differences between modalities, particularly in heterogeneous images. How to deal with modal differences while effectively integrating change information remains a challenge in MCD. In this article, we propose a novel adversarial feature equilibrium network (AFENet), which establishes an additional adversarial optimization to solve the equilibrium problem between modal differences and land cover changes. Our AFENet aligns the features and reduces the modal gap through a multiscale adversarial domain adaptation (MADA) approach. Meanwhile, a divergence-aware contrastive module (DCM) is designed as a regularization term for adversarial optimization. DCM affects the sensitivity of feature extractors by constraining the mutual information between changed and unchanged pixels. In this case, AFENet can maintain the consistency of feature representation while maximizing the discriminability of change targets. The features extracted from AFENet will then be integrated by our multistream feature fusion (MFF) module and utilized to generate change maps. The effectiveness of our approach is demonstrated on two scene-level multimodal RS datasets. Compared with existing methods, our AFENet achieves state-of-the-art (SOTA) performance on both datasets and outperforms the second-best$F1$score by 4.64% and 1.1%, respectively.

Remote sensing scene classification (RSSC) aims to understand and analyze the semantic information at the scene level with complex geographical properties. Despite the profound success of advanced deep models in automatically capturing hierarchical embedding representations and the gradual dominant trend in RSSC, it still remains a great challenge to precisely focus on targets at variable scales that are considered highly relevant to the corresponding scene and separated from the background. Motivated by this recognition, in this article, we present the gradient-guided multiscale focal attention network (GMFANet) for RSSC to adaptively localize the representative multiscale semantic representation for complex scenes. In particular, a lightweight parameterized hierarchical multiscale attention (HMA) mechanism is proposed, which constitutes the main aim of adaptively enhancing physical detail and high-level semantic information at different layers, rather than regarding each scale set with equivalent insight, while eliminating redundant information inherent in conventional attention mechanisms. Subsequently, a gradient-guided spatial focused attention (GSFA) module is specifically designed to accurately localize critical regions at multiple scales, with the dynamic combination of gradient-activated reference attention map and prediction attention map from supervised information-based learning. In addition, a curriculum-driven dynamic attention fusion (CDAF) strategy is tailored to fuse the spatial attention above from easy to hard for avoiding from poor local optimum and decreasing the early learning ambiguity. Our extensive comparative experiments and ablation analyses implemented on real-world public RSSC datasets indicate that our approach achieves the state-of-the-art performance exactly. The code is available athttps://github.com/bling2beyond/GMFANet.

With the rapid development of remote sensing (RS) technology, high-resolution RS image change detection (CD) has been widely used in many applications. Pixel-based CD techniques are maneuverable and widely used, but vulnerable to noise interference. Object-based CD techniques can effectively utilize the abundant spectrum, texture, shape, and spatial information but easy-to-ignore details of RS images. How to combine the advantages of pixel-based methods and object-based methods remains a challenging problem. Besides, although supervised methods have the capability to learn from data, the true labels representing changed information of RS images are often hard to obtain. To address these issues, this article proposes a novel semisupervised CD framework for high-resolution RS images, which employs small amounts of true labeled data and a lot of unlabeled data to train the CD network. A bihierarchical feature aggregation and extraction network (BFAEN) is designed to achieve the pixelwise together with objectwise feature concatenation feature representation for the comprehensive utilization of the two-level features. In order to alleviate the coarseness and insufficiency of labeled samples, a confident learning algorithm is used to eliminate noisy labels and a novel loss function is designed for training the model using true- and pseudo-labels in a semisupervised fashion. Experimental results on real datasets demonstrate the effectiveness and superiority of the proposed method.

Change detection (CD) in synthetic aperture radar (SAR) images aims to detect changed areas by considering the changes in backscattering coefficients. However, the changes can be further divided into positive and negative changes in terms of the increase or decrease of backscattering coefficient, so the CD task can be divided into binary and ternary according to the number of existent categories. This paper introduces an M-nary (binary or ternary) SAR change detection procedure based on the generative adversarial network (GAN) and neural architecture search (NAS) strategy to detect which changes exist in the SAR image-pair and design specialized classifiers for both binary and ternary CD. First, a difference image generation approach based on the salient changed region extraction and neighborhood information is designed for a robust difference representation on the M-nary CD. Due to the further subdivision of changes, the insufficiency of labeled data presents the M-nary change detection with a dilemma. Concerning the lack of labeled information, this paper presents a labeled sample generation strategy based on the GAN architecture search to supplement sample data. Since GAN training is inherently unstable, NAS provides an effective means of searching GAN architecture automatically and ameliorates the reliability of generated samples. During the architecture search procedure, a double-phase evolutionary search strategy is introduced to further improve the stability of GAN training. The experimental results with theoretical analysis prove the validity, robustness, and potential of our method in synthetic as well as real SAR datasets.

Classification of remote sensing image series which differ in quality and details, has impportant implications for the analysis of land cover, whereas it is expensive and time-consuming as a result of manual annotations. Fortunately, domain adaptation (DA) provides an outstanding solution to the problem. However, information loss while aligning two distributions often exists in traditional DA methods, which impacts the effect of classification with DA. To alleviate this issue, an inter-domain representation alignment and fine-tuning based network (RAFNet) is proposed for image series classification. Inter-domain representation alignment, which is fulfilled by a variational auto-encoder (VAE) trained by both source and target data, encourages reducing the discrepancy between the two marginal distributions of different domains and simultaneously preserving more data properties. As a result, RAFNet, which fuses the multi-scale aligned representations, performs classification task in the target domain after well trained with supervised learning in the source domain. Specifically, the multi-scale aligned representations of RAFNet is acquired by duplicating the frozen encoder of VAE. Then, an information based loss function is designed to fine-tune RAFNet, in which both the unchanged and changed information implied in change maps is completely used to learn the discriminative features better and make the model more generalized for the target domain. Finally, experiment studies on three datasets validate the effectiveness of RAFNet with considerable segmentation accuracy even the target data has no access to any annotated information.

In this paper, we propose a new change detection method of multi-temporal synthetic aperture radar (SAR) images. Due to the ability of extracting key feature of images and robustness to noise, stacked denoising auto encoder (SDAE) has been widely used in remote sensing. However, the single SDAE stills has some limitations to handle with the speckle noise of SAR images. Therefore, we propose a new structure Bi-hierarchical SDAE for feature extraction. The first level of SDAE denoises the original image and reconstructs the difference map, and the second level extracts the superpixel-based difference features for classification. Besides, Bayesian optimization effectively improves the classification performance of feature classifier. The experimental results of the datasets in this paper show that the Bi-hierarchical SDAE and Bayesian optimization framework has high accuracy and proves its effectiveness.

Abstract As 21st‐century skills have become increasingly important, collaboration ability is now considered essential in many areas of life. Different theoretical frameworks and assessment tools have emerged to measure this skill. However, more applied studies on its implementation and assessment in current educational settings are required. This research accordingly uses Graph Convolutional Neural Networks (GCNs) to assess students' collaboration ability from students' assignments. The Pearson correlation coefficient is used to measure the similarity of the level of students' collaboration ability, and similar students are linked together to establish an adjacency matrix. By sorting through relevant literature and selecting the feature words that represent the strength of collaboration ability, calculating the similarity between the preprocessed student data and each selected feature word, after which the highest value of the similarity as the feature value of the student for this feature and establish the student feature matrix. Finally, the GCNs are jointly trained by the adjacency matrix and the feature matrix. The results show that this method can effectively assess students' collaboration ability. Moreover, compared with other text classification methods, the GCNs selected in this paper has higher accuracy.

Emerging memory devices are an attractive choice for implementing very energy-efficient in-situ matrix-vector multiplication (MVM) for use in intelligent edge platforms. Despite their great potential, device-level non-idealities have a large impact on the application-level accuracy of deep neural network (DNN) inference. We introduce a low-density parity-check code (LDPC) based approach to correct non-ideality induced errors encountered during in-situ MVM. We first encode the weights using error correcting codes (ECC), perform MVM on the encoded weights, and then decode the result after in-situ MVM. We show that partial encoding of weights can maintain DNN inference accuracy while minimizing the overhead of LDPC decoding. Within two iterations, our ECC method recovers 60% of the accuracy in MVM computations when 5% of underlying computations are error-prone. Compared to an alternative ECC method which uses arithmetic codes, using LDPC improves AlexNet classification accuracy by 0.8% at iso-energy. Similarly, at iso-energy, we demonstrate an improvement in CIFAR-10 classification accuracy of 54% with VGG-11 when compared to a strategy that uses 2× redundancy in weights. Further design space explorations demonstrate that we can leverage the resilience endowed by ECC to improve energy efficiency (by reducing operating voltage). A 3.3× energy efficiency improvement in DNN inference on CIFAR-10 dataset with VGG-11 is achieved at iso-accuracy.

Graph Cuts is a popular technique for Maximum A Posteriori inference in computer vision. It transforms a Markov Random Field problem into a network flow problem, solved via the Push-Relabel algorithm. While attractively simple, the large size of a typical image and the large number of necessary pixel-level iterations render the technique computationally expensive. Prior accelerator attempts have been reported with GPUs and FPGAs. In 2017, we demonstrated the first pixel-parallel architecture on FPGA, but limited to only 256-pixel images. This paper extends this pixel-parallel concept and proposed a Virtual-Image architecture which solves the size limitation. We demonstrate the first working virtual-image Graph Cuts accelerator, implemented on a state of the art FPGA, applied to standard benchmark images for a background segmentation task. The design is 11-13× faster than other FPGA designs, and slightly faster than a modern GPU benchmark by about 30%.

A Pixel-Parallel Virtual-Image Architecture for High Performance and Power Efficient Graph Cuts Inference Tianqi Gao, University of Illinois Urbana Champaign Rob A. Rutenbar, University of Pittsburgh Contact: [email protected] Graph Cuts is a popular technique for Maximum A Posteriori inference in computer vision. It transforms a Markov Random Field problem into a network flow problem, solved via the Push-Relabel algorithm. While attractively simple, the large size of a typical image and the large number of necessary pixel-level iterations render the technique computationally expensive. Prior accelerator attempts have been reported with GPUs and FPGAs. In [1], the first pixel-parallel architecture was demonstrated in FPGA, but limited to only 256-pixel images. This paper extends this pixel-parallel concept and makes following contributions: a "Virtual Image" architecture solves the size limitation: large images are decomposed into "tiles", and "stacked" on the physical processor array; appropriate addressing mechanisms handle virtual pixels and a range of tile edge effects; scaling up the processor array to 1536 pixels; a novel and hardware-friendly heuristic shortens the convergence. We demonstrate the first working virtual-image Graph Cuts accelerator, applied to standard 640x480 images. Scaling up the hardware and the new heuristic bring 6.5x and 1.65x speedups respectively compared with [1]. The design is 7-20x faster than prior FPGA designs, and roughly comparable in speed to a modern GPU benchmark. However, the architecture also offers significant performance-per-unit-power advantages. Formulating a figure of merit particularly for Graph Cuts inference - Graph Cuts per second per Watt - our architecture is about 4 times better than other implementations. Keywords: FPGA; Machine learning; Hardware Acceleration; Computer Vision DOI: https://doi.org/10.1145/3289602.3293948

In this work, a spin torque nano oscillator (STNO) device whose frequency can be tuned by Voltage Controlled Magnetic Anisotropy effect (VCMA) is proposed. The requirement of magnetic bias field in previous STNO devices is eliminated by the introduction of VCMA effect. Based on VCMA-STNO, a novel architecture is proposed which can compose of a modulation/demodulation digital frequency shift keying (DFSK) communication system. The proposed architecture utilizes VCMA-STNO as core devices and is much simpler comparing with its CMOS counterpart. The proposed VCMA-STNO modulation/demodulation architecture will help to design next generation spintronics DFSK communication system.

The method of Graph Cuts converts a Maximum a Posteriori (MAP) inference problem on a Markov Random Field (MRF) into a network flow, which can be solved efficiently. Many computer vision problems can be conveniently cast as an inference task to find most likely labels for pixels. The method is widely used, but computationally burdensome. Prior accelerator attempts have failed to exploit the problem's attractive, maximum available parallelism: push-relabel flow solvers can run in parallel across every pixel. This paper describes the design and implementation of the first pixel-parallel Graph Cuts inference engine. Our prototype implements a 256-pixel tile of an image, implemented as 256 locally-connected pixel processors. A checkerboard scheduling scheme allows for maximum parallelism while avoiding critical data dependencies. A 150MHz implementation on an FPGA can solve a segmentation task in 6 microseconds. We also discuss strategies for extending our prototype to larger "virtual" images that span more than the physical extent of the inference tile. Our model suggests 2-40× speedups compared with previous accelerator experiments. To the best of our knowledge, this is the first fully functional, pixelparallel accelerator demonstration for Graph Cuts inference.

It is well known that, by properly selecting the measurement configurations in robot calibrations, the observability index of unknown parameters can be maximized, leading to high calibration accuracy. For this purpose, many configuration-search methods were proposed. However, the established methods were mainly based on derivative-free or metaheuristic techniques, whose computational costs were high. Moreover, the robustness of observability index and convergences of configuration searches were not investigated. In this paper, by extending a recent result in matrix perturbation theory to robot kinematics, we establish the closed-form mapping from configuration perturbations to singular-value variations. Based on this mapping, an efficient configuration-search method is proposed, the robustness of the observability index under bounded configuration perturbations is analyzed, and the convergence of configuration searches is studied. The proposed methods were validated by simulations on serial and parallel robots. With roughly estimated initial parameters, self-calibration experiments on a redundant cable-driven parallel robot were performed. The effectiveness of the proposed methods is demonstrated by the experiment results.

In this work, we have proposed that the neural synapses and neurons can be realized by utilizing spin waves (SWs) as information carrier. The SWs is excited by spin torque nano-oscillator (STNO), and detected with several different physical mechanisms: 1) tunneling magnetic-resistance 2) spin pumping and 3) inverse spin hall effect. The proposed SWs based synapses and neurons can be further combined together to form a neuromorphic computation system with crossbar structure. Possible ultra low power consumption and ultra high speed are the advantage of our proposed SWs based synapses and neurons.