Spiking Neural Networks (SNNs) have emerged as a promising paradigm for brain-inspired edge computing. Leveraging binary spikes and local learning rules, SNNs enable energy-efficient on-chip learning and rapid adaptation to changing environments, which is crucial for edge AI that needs to learn continuously from new data. However, many SNN processors enabling on-chip learning for edge computing confront a trade-off: small-scale task-specific designs offer low power but poor multi-task inference accuracy, while large-scale general-purpose designs achieve high multi-task accuracy at the cost of large memory and poor energy efficiency. To overcome this challenge, this paper presents ANP-R, a 22nm asynchronous SNN-based edge AI processor with coarse-grained reconfigurable architecture enabling one-shot, few-shot, batch and incremental on-chip learning. The processor integrates 64 cores containing 4096 neurons and 0.262 million synapses. Two key features are proposed: 1) An asynchronous coarse-grained reconfigurable architecture that supports various STDP-based SNN topologies. These topologies enable over 95% average accuracy across four sensory tasks; 2) an energy-efficient asynchronous training method incorporating a self-adaptive synaptic weight update mechanism reducing up to 65% redundant updates without accuracy loss, and a trained weights low-bit width coding method reducing up to 50% storage cost with 0.3% accuracy loss. Measurement results demonstrate 92.1% accuracy for hand gesture classification, 93.9% for keyword spotting, 98.6% for object recognition and 99.2% for gas identification. Compared with state-of-the-art SNN-based chips, this work achieves up to 6.02x, 8.61x and 7.1% improvement in energy efficiency, energy per step, and accuracy, respectively.
Recent contrastive multi-view clustering methods have achieved remarkable performance by using two-branch contrastive learning. However, most existing studies focus on the optimization of the false negatives (FNs) identification strategy, ignoring the critical issue of cluster center alignment between fused view and single views. To address this limitation, we present a robust multi-view clustering method (CAFE) based on cross-view adaptive fusion and cluster center enhancement. Specifically, we first design a cross-view adaptive fusion module that incorporates dual weights at both the view level and the sample level, enabling effective coordination of consistent and complementary information across views. Subsequently, we propose a dual-driven cluster center enhancement framework to refine cluster structures. It introduces a dual alignment mechanism between single-view cluster centers and fused-view cluster centers to systematically coordinate view-specific discriminative patterns and cross-view consensus representations. Furthermore, we develop a second-order proximity graph embedding method to more effectively rectify FNs by computing neighborhood similarity. It constructs second-order proximity to identify structurally related samples that may be spatially distant in feature space. Extensive experiments on six widely used multi-view benchmark datasets demonstrate that CAFE achieves state-of-the-art performance under both complete and incomplete multi-view scenarios.
This article proposes a quantum self-attention neural network (QSAN) model that is implementable on parameterized quantum circuits (PQCs), providing a novel avenue for addressing natural language processing (NLP) tasks via quantum computing. The QSAN architecture comprises four blocks: the data preprocessing block, the quantum encoding block, the model design block, and the network optimization block. Through these blocks, classical text data are initially preprocessed and encoded into quantum states. The rich semantic features and intricate relationships within these quantum states are then captured and learned by a quantum self-attention layer and a quantum fully connected layer. Finally, the model performance is enhanced by optimizing the network parameters. Simulation results on publicly available topic classification and sentiment analysis datasets demonstrate that the QSAN model outperforms state-of-the-art baselines. As a limitation of the current study, these evaluations are restricted to a maximum sentence length of eight words due to simulation and hardware constraints. Moreover, the effectiveness, scalability, trainability, and robustness of the proposed model are comprehensively evaluated, highlighting its superior performance and potential for advanced NLP tasks.
Technological advances have enabled new ways for individuals with severe motor impairments (SMI) to interact with computer-based applications, yet standard human-computer interfaces often remain inaccessible, creating a need for assistive technologies (AT) tailored to diverse motor abilities. This scoping review aims to identify existing AT solutions that enable individuals with SMI to use residual movements for controlling digital applications, characterizing key trends and gaps. Searches were conducted in Scopus and Web of Science Core Collection for studies published between 2012 and June 2024, and data collection took place from July to December 2024. In total 49 articles were included in the review. In addition to descriptive indicators, cross-topic analyses were performed using relationship and correspondence analysis. Overall, the AT landscape for SMI users is dominated by head-based control and 2D pointer interaction for tasks such as typing and navigation. Three-quarters of the publications addressed SMI generically, without specifying underlying conditions or detailed movement profiles. Vision systems dominate, followed by electromyographic and kinematic setups. Most systems employed explicit direct mappings, with only 20% using machine learning. A quarter of systems incorporated auxiliary feedback, with haptics used in only a small minority. Despite advancements in sensor technology and signal processing, improvements are still needed in key areas: integrating multimodal input and feedback, applying machine learning to the design process, and clearly defining target motor profiles. Future research should address these gaps, exploring novel use cases beyond conventional computing tasks and expanding participation among individuals with SMI. The scarcity of studies defining or targeting specific motor profiles underscores the importance of rehabilitation professionals advocating its critical importance in tailoring interventions that connect movement capabilities with clear, goal-oriented digital activities, thus fulfilling the Human Activity Assistive Technology (HAAT) model’s emphasis on activity goals.The findings highlight a lack of studies investigating the multimodal input and feedback systems that facilitate effective interaction with digital tasks, underscoring the need for increased research in this field.Machine learning approaches in assistive technology design enhance adaptability and personalisation, enabling devices to better interpret and respond to users’ unique motor profiles and activity goals. The limited incorporation of machine learning approaches in AT design suggests that rehabilitation professionals should actively contribute to a wider demand for its application.Increasing participation of individuals with SMI in the evaluation of AT systems is essential for developing inclusive solutions that address diverse needs and enhance real-world digital engagement.
Superabsorbent polymer (SAP) microparticles incorporating 2-hydroxymethyl-12-crown-4 were developed to combine Li+-selective coordination with rapid hydrogel swelling. The materials were prepared in the presence of Li+-crown ether complexes within an acrylic acid-acrylamide (AA-AAm) network, resulting in coordination environments within the swollen hydrogel structure. Single-particle swelling measurements showed that the microparticles reached equilibrium within ∼5 min, enabling rapid solution ingress and transport of Li+ into the hydrogel network. In batch adsorption experiments, the modified SAPs achieved a lithium adsorption capacity of ∼1200 μmol g-1 at 20 ppm Li+ (pH 7), significantly higher than non-functionalized SAPs and a rigid methacrylic acid-ethylene glycol dimethacrylate (MAA-EGDMA)-based IIP included as a literature-class benchmark for field contextualization. The materials also exhibited preferential Li + uptake over Mg2+ and Co2+ under both single-ion and competitive conditions and retained higher adsorption capacity under acidic conditions (pH 4.5). These results suggest that incorporation of crown ether functionality within a fast-swelling hydrogel matrix enhances Li+ uptake and selectivity through a combined effect of ion transport and coordination interactions.
Glucagon-like peptide-1 receptor agonists (GLP-1RAs) have reshaped the clinical approach to managing obesity and type 2 diabetes. As approved indications have expanded, use of GLP-1RAs has increased rapidly in the United States. Although randomized trials demonstrate strong efficacy, many questions remain about their optimal use in clinical practice. Real-world data (RWD) from electronic health records, registries, insurance claims, and other sources offer a promising avenue to address these questions. However, concerns about data quality, selection bias, and incomplete ascertainment of medication use and outcomes pose significant challenges to the validity of the resulting evidence. In May 2025, the National Institute of Diabetes and Digestive and Kidney Diseases convened experts from regulatory agencies, payer organizations, and academia to explore these challenges. This second of 2 synopsis articles on the workshop summarizes the discussion around the strengths and limitations of various RWD sources and methodological approaches to strengthen causal inference and generalizability. Presenters highlighted pragmatic clinical trials and target trial emulation as strategies to generate stronger real-world evidence (RWE) that is relevant to both clinical practice and policy. The workshop underscored that careful attention to study design, data limitations, and analytic approach is essential to yield RWE that informs clinicians, patients, payers, and policymakers.
COVID-19 is a highly contagious disease transmitted primarily through human contact. Therefore, understanding population mobility is essential for predicting COVID-19 case trends. In this paper, we propose a novel deep learning approach for forecasting new COVID-19 cases using a neural architecture called Neural Basis Expansion Analysis for Interpretable Time Series (N-BEATS). The N-BEATS model effectively handles long input sequences and large output horizons without information loss or increased computational complexity. We compare the performance of N-BEATS with a state-of-the-art benchmark model, LSTM-Markov, across four major countries: the United States, the United Kingdom, Russia, and Brazil. Three distinct COVID-19 datasets from Google, Apple, and Our World in Data (OWID) were used in this study. Incorporating Google and Apple mobility data as covariates enhances both the accuracy and interpretability of the N-BEATS model. Our results show that N-BEATS consistently outperforms LSTM-Markov across all datasets and countries, consistently yielding lower Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). Furthermore, the N-BEATS model with covariates outperforms its counterpart without covariates, indicating that mobility data provide substantial value for forecasting new COVID-19 cases. Overall, this study demonstrates the effectiveness of the N-BEATS architecture in capturing pandemic dynamics and offers valuable insights for policymakers and public health officials in managing future outbreaks.
Visual perspective-taking requires mentally adopting another spatial viewpoint by aligning one's own perspective with that of a reference, such as another person or object. This alignment process is thought to rely on embodied spatial transformations engaging both spatial computation and motor-related systems. Using functional magnetic resonance imaging (fMRI), we tested whether neural responses scale with angular disparity between observer and reference viewpoints. Participants judged the relative position of objects from the perspective of either an avatar or an empty chair across alignment conditions and rotation angles. Unaligned conditions elicited stronger activation than aligned conditions in a distributed network encompassing medial frontal, premotor, opercular, posterior parietal, ventral temporal, and cerebellar regions. Both reaction time and Blood-Oxygen-Level-Dependent signal (BOLD) magnitude increased linearly with angular disparity, and individual differences in response time correlated with activation magnitude, linking behavioral effort to neural computation. Comparable results were obtained when the reference was a chair, indicating that these mechanisms generalize beyond social contexts. Together, these findings reveal a cerebello-frontoparietal network that supports perspective alignment through embodied simulation, spatial transformation, and predictive modeling, providing a unified account of the neural computations underlying VPT.
Brassinosteroids (BR) are an essential steroid phytohormone found throughout the plant kingdom. Synthesis of BR occurs at the endoplasmic reticulum, and the mechanism of transport through the membrane to the apoplast is not well understood. The ABCB1 and ABCB19 transporters enhance BR transport through the plasma membrane, but the available cryo-EM structures do not provide a clear understanding of the steroid's diffusion process. Sphingolipid and sterol enriched domains improve the translocation of BR through localization of ABCB1 and ABCB19. The similarity of BR to sterols suggests that BR localizes differently between the more disordered inner leaflet and the tightly packed outer leaflet which contains sphingolipid and sterol domains. The presented work utilizes all-atom molecular dynamics (MD) simulations to compare the diffusion properties of BR with β-sitosterol. Our results show key thermodynamic differences between BR and sitosterol passive transport and a pathway to exposure of BR to ABCB entry sites.
Endogenous pain modulation is believed to encompass a crucial evolutionary purpose in guiding decision-making away from harm. This is well exemplified by the idea that mechanisms of learning through error correction and endogenous pain modulation are inherently intertwined. However, adaptive behavior requires more than learning through error correction. Biological environments are volatile, which can cause decision-makers to be uncertain about what actions lead to rewards or punishments. Evidence on how uncertainty in action-outcome distributions impacts endogenous pain modulation is lacking. In this study, we extend and adapt a well-established paradigm for the study of endogenous pain modulation with the implementation of a reversal learning outcome schedule. Thirty healthy human volunteers took part in this probabilistic gambling task, where they had to gamble for the obtainment of pain relief and the avoidance of pain punishments. Using a computational approach, we assess mechanisms of uncertainty in learning in acute pain situations. Such uncertainty is associated with the observed pain modulation in the task, indicating that endogenous pain modulation may be sensitive to volatility and the perception of uncertainty. Specifically, pain inhibition from winning pain relief increases as a function of certainty in what actions lead to pain relief. Our findings emphasize the importance of considering mechanisms of uncertainty processing in reinforcement learning from painful outcomes and endogenous pain modulation. Those mechanisms could be relevant to understanding behavioral changes in chronic pain, where altered reinforcement learning has already been established.
Estimating the number of insect species on Earth is a daunting challenge. The current consensus estimate-about six million species-is likely far too low, as we will show. Our estimate of the global number of insect species rests on a sample of more than 1,600,000 DNA-barcoded insect specimens representing 53,945 species from 15 "core" Malaise traps deployed in dry forest, cloud forest, and rainforest ecosystems of the Área de Conservación Guanacaste (ACG) in Costa Rica. Even this massive sample fails to reveal the full extent of ACG insect species richness. To estimate total ACG insect richness, we adjust the observed count of insect species by an "undersampling ratio," computed for a hyperdiverse subfamily of parasitoid wasps (Braconidae: Microgastrinae). The ratio compares microgastrine richness from the core Malaise traps to a lower-bound estimate of true microgastrine richness-including undetected species-based on 21,669 specimens from three sources: the 15 core Malaise traps, 15 "peripheral" Malaise traps spanning all three ecosystems, and 11,373 DNA-barcoded specimens reared from some 1,500 species of microgastrine-parasitized caterpillars (Lepidoptera). To estimate global insect richness, we apply Earth/ACG ratios for tree species and several animal taxa to upscale our estimate of ACG insect richness (nearly 333,000 species). Adopting conservative assumptions, we reach an estimate of 14 to 20 million insect species on Earth, depending on the upscaling group-two to three times the current consensus estimates. Upscaling instead from a point estimate of ACG richness with a wide CI, global estimates reach nearly 30 million species.
Deciphering protein function is fundamental to advancements in medicine and biotechnology. However, conventional experimental characterization remains resource-intensive. Public large language models (LLMs), though proficient in natural language processing, often fail to accurately interpret and predict the functional and structural properties of proteins, limiting their utility in bioinformatics. To address this gap, we introduce BetaDescribe, designed to generate detailed and rich textual descriptions of proteins, including their function, catalytic activity, involvement in specific metabolic pathways, subcellular localizations, and the presence of specific domains. The trained BetaDescribe model receives protein sequences as input and outputs a textual description of these properties. BetaDescribe starting point was the LLAMA2 model, which was trained on trillions of tokens. Our model was next trained on datasets containing both biological and English text, which allowed the incorporation of biological knowledge. In addition to the description generator, BetaDescribe comprises multiple validator models and a judge, which together enable accurate ranking of alternative generated descriptions. We demonstrate the utility of BetaDescribe by providing descriptions for proteins that share little to no sequence similarity to proteins with functional descriptions in public datasets. Using in silico mutagenesis, we further show that BetaDescribe relies on functionally important regions, as part of its prediction, suggesting that the model identifies regions of importance for the protein functionality without needing homologous sequence. BetaDescribe offers a powerful tool to explore protein functionality, augmenting existing approaches such as annotation transfer based on sequence or structure similarity.
In this study, artificial root dentine lesions were treated with a novel peptide, GA-C16G2, to evaluate its remineralisation potential under in vitro conditions. Human root dentine specimens were demineralised for 4 days and subjected to a 7-day pH-cycling regimen with daily application of GA-C16G2 (500 μM), C16G2 (500 μM), gallic acid (GA; 500 μM), or distilled water (DW). Mineral density, mechanical properties, morphology, and chemical composition were determined using micro-computed tomography, nanoindentation, scanning electron microscopy, energy dispersive spectroscopy, and Fourier transform infrared spectroscopy analyses, respectively. GA-C16G2 and GA treatments resulted in significantly higher mineral density (1.69 ± 0.05 g/cm3 and 1.69 ± 0.02 g/cm3) than C16G2 (1.58 ± 0.06 g/cm3) and DW (1.63 ± 0.04 g/cm3) (p < 0.05), corresponding to mineral gains of 17.8% and 16.5%, respectively. Similarly, GA-C16G2 (hardness: 0.33 ± 0.08 GPa; elastic modulus: 7.63 ± 1.93 GPa) and GA (hardness: 0.32 ± 0.06 GPa; elastic modulus: 6.91 ± 1.45 GPa) demonstrated significantly greater recovery of mechanical properties than C16G2 and DW (p < 0.05). Root dentine appeared to show fewer exposed collagen fibrils, higher calcium and phosphate contents, and stronger phosphate bands in groups GA-C16G2 and GA, compared to those in groups C16G2 and DW. GA-C16G2 demonstrated the potential to promote mineral recovery and mechanical reinforcement of artificial root dentine lesions. Further studies incorporating established remineralising reference standards and biofilm-based models are required to determine its relative therapeutic performance and clinical relevance.
暂无摘要（点击查看详情）
Hippocampal formation (HF) supports both the temporary maintenance of task-relevant information and rapid relearning when task structure is preserved. Here we ask what circuit mechanism can link these two functions within a single framework. We propose a model named Generalization and Associative Temporary Encoding (GATE), whose core idea is a self-gating re-entrant EC3-CA1-EC5-EC3 loop. In each lamella, EC3 provides a memory substrate, CA1 selectively reads out the retained information under CA3 gating, and EC5 feeds back to regulate the next EC3 state. Repeating this loop across dorsoventral lamellae yields representational scales that range from local cue-dependent coding to a broader task-related structure. In simple tasks, the single-lamellar model captures selective maintenance and produces place- and splitter-like CA1 activity. In more complex tasks, the multi-lamellar model develops lap, evidence, trace, and other task-relevant representations. Under structure-preserving changes in sensory coding, positional scaffold, or task parameters, the model reuses learned representations and relearns faster. GATE provides a hypothesis-generating computational framework for studying how hippocampal-like circuit motifs may support selective memory gating and structure-preserving relearning.
Neural stability is essential for executing learned motor behaviors while plasticity provides the flexibility needed to adapt to new tasks and environments. Although low-dimensional neural population dynamics exhibit long-term stability, the extent to which individual neurons retain their functional properties over time and balance the need for both stability and plasticity remains an open question. Tracking individual neurons across multiple recording sessions is crucial to addressing this question, yet conventional methods face challenges such as electrode drift, waveform variability, and large inter-electrode distances that limit the number of channels a neuron is observed on. Here, we introduce a waveform-based neuron tracking method optimized for standard microelectrode arrays, enabling the identification of the same neurons across sessions without relying on spatial overlap, a strategy commonly leveraged with high-density electrode arrays. We apply this method to assess the longitudinal stability of multiple neural properties, including firing rates, inter-spike intervals, tuning properties, and spike-field interactions. Our findings reveal that while spike waveform properties remain stable, certain functional properties such as ISI and tuning can exhibit gradual shifts, suggesting a balance between neural stability and plasticity. Understanding the persistence of individual neural signals provides insight into learning and adaptation while advancing the study of neural stability and plasticity over extended timescales. Beyond basic neuroscience, this framework has potential to enhance the long-term reliability of brain-machine interfaces and closed-loop deep brain stimulation systems that rely on chronic neural sensing.
The acquisition and prediction of mechanical information were important for the digitization of Chinese Traditional Mongolian Osteopathy. The extraction of periodic and trend-oriented mechanical information from the force exertion process presented a substantial challenge in achieving precise mechanical information prediction. To tackle this challenge, this article introduced a double-layer Long Short-Term Memory (LSTM) network integrated with Hodrick-Prescott (HP) filtering, which combined HP filtering techniques with LSTM to enhance the accuracy of mechanical information prediction. This research was grounded in experimentally collected data for mechanical information analysis, wherein dimensionality reduction was performed on the acquired data based on the primary joints involved in force application. Specifically, three sensor data points were selected from a total of 24 as pivotal input features. LSTM was employed to capture long-term dependencies inherent in sequential data. In this investigation, the Adam optimization algorithm was utilized to fine-tune the model's hyperparameters, ensuring optimal performance. The experimental outcomes underscored the efficacy of the proposed methodology, evidenced by a coefficient of determination (R2) of 0.897, a mean absolute error of 0.025, and a root mean square error of 0.031. A comparative analysis with alternative estimation methods further attested to the stability, accuracy, and generalization capabilities of the two-layer LSTM network augmented with HP filtering.
From hunting to financial planning, many decisions are made under risk and alongside social partners whose decisions and outcomes are coupled with our own. Joint decision-making can create challenges, such as navigating conflicting risk preferences, evaluating each other's actions, and deciding whether to compromise. Yet when and why people compromise, how they evaluate each other's responsibilities during joint decision-making, and the computational and psychological underpinnings of these processes remain unclear. We introduce a dyadic foraging paradigm designed to capture diverse risk preferences, where two participants jointly choose between locations that yield rewards but carry risks and evaluate each other's responsibility for shared outcomes. Across two studies (exploratory N = 250, confirmatory N = 514), people tended to compromise rather than counteract their partner, especially under diverging risk preferences and when compromise was reciprocated. Computationally, compromise was explained by a reinforcement learning model showing that individuals integrate their preferences with that of their partner. Responsibility attributions exhibited egocentric biases-with participants claiming more credit for wins than blame for losses-and these biases were associated with individual differences in compromise behavior. The interplay between individual differences in risk and responsibility attribution further shaped coordination patterns. Finally, we show that compromising improved performance for risk-averse individuals, increased desirability as a social partner, and led to more favorable responsibility attributions, suggesting multiple benefits of compromising. By linking decisions about risk-reward trade-offs to metacognitive judgments about responsibility, our study reveals the social and cognitive processes underlying compromise in risky foraging and, more broadly, in collaborative contexts with conflicting preferences.
Identifying causal genetic variants in a computational manner remains an open problem. Training end-to-end prediction models is not possible without large ground-truth datasets, while results of genome-wide association studies (GWAS) are entangled by linkage disequilibrium (LD), and gene expression datasets do not contain genetic variation at individual-level. Here, we propose Multiple Instance Fine-mapping (MIFM) - a multiple instance learning (MIL) objective to overcome the lack of strong labels by grouping putatively causal variants together based on their LD scores. Using MIFM, we trained a deep classifier on a dataset aggregating over 13,000 GWAS to predict causal variants based on their underlying DNA sequences. We validated variants prioritized by MIFM by constructing polygenic risk scores which transferred better to different target ancestries. Furthermore, we demonstrated how MIFM can be used to disentangle effect sizes of highly-correlated variants to better fine-map GWAS results.
Brain activity associated with bilateral hand movements has been reported to involve multiple regions. However, the association between gray matter (GM) volume in these regions and bilateral hand movement remains unclear. Therefore, this study aimed to elucidate the association between tapping skills using both hands and GM volume. This study included 61 healthy adults. Each participant's Magnetic Resonance Imaging scans and bilateral motor task performance were evaluated. The motor task was a finger-tapping task, and the movement conditions were set as in-phase and antiphase. The participants were instructed to open and close their thumb and index finger as quickly and widely as possible, either simultaneously (in-phase condition) or alternately (antiphase condition) on both sides. Using voxel-based morphometry, we measured brain GM volume associated with the performance in the bilateral finger-tapping task. The results indicated a significant negative correlation between the similarity index of bilateral movement under the antiphase condition and GM volume around the left cingulate motor cortex. This result suggests that participants who were able to smoothly move both hands under the antiphase condition had greater GM volume in the left cingulate motor cortex. This GM volume may serve as a neurophysiological indicator of bilateral hand movement.