187037 research outputs found
Sort by
Teach sample-specific knowledge: Separated distillation based on samples
No full textRecent advancements in deep neural networks have revolutionized computer vision, enabling practical applications like classification and object detection. However, deploying these models on resource-constrained devices remains a critical challenge due to their high computational demands. Knowledge Distillation (KD) has emerged as an effective technique to address this issue by transferring knowledge from complex teacher models to lightweight student models, enhancing efficiency while maintaining high performance. Traditional logit-based KD methods use forward Kullback-Leibler divergence (FKLD) to transfer meaningful knowledge. However, FKLD typically exhibits a mode-averaging property, causing students to focus on non-target information, whether the teacher's samples are correct or incorrect. Additionally, when handling uncertain samples, even teacher models may fail to classify them accurately, leading to incorrect predictions and confusing the students. To address these issues, we classify the dataset into two groups based on the teacher's predictions: correct and incorrect samples. To ensure a more reliable transfer of knowledge from teacher to student for correct samples, we employ both forward Kullback-Leibler divergence (FKLD) and reverse Kullback-Leibler divergence (RKLD), which has mode-focusing properties. We also reduce temperature scaling for RKLD to enhance the focus on target information, ensuring that the student model prioritizes meaningful knowledge while minimizing the influence of non-target information. Conversely, for incorrect predictions, our method minimizes the teacher's knowledge, encouraging students to rely more on the true labels by focusing on cross-entropy loss. Experimental results on both classification and object detection tasks demonstrate that our method, Teach Sample-Specific Knowledge (TSSK), outperforms state-of-the-art KD methods, making it ideal for deployment on-devices in real-world scenarios.
ELiOT: End-to-end LiDAR odometry with transformers harnessing real-world, simulated, and digital twin
No full textThe development of smart cities depends on intelligent systems that integrate data from diverse environments. In this work, we present ELiOT, an end-to-end LiDAR odometry framework with transformer architecture designed to utilize real-world data, simulations, and digital twins. ELiOT leverages high-fidelity simulators and digital twin environments to enable sim-to-real applications, training on the real-world KITTI odometry dataset while benefiting from simulated data for improved generalization. Our self-attention-based flow embedding network eliminates the need for traditional 3D-2D projections by implicitly modeling motion from sequential LiDAR scans. The framework incorporates a 3D transformer encoder-decoder to extract rich geometric and semantic features. By integrating digital twin environments and simulated data into the training process, ELiOT bridges the gap between simulation and real-world applications, offering robust and scalable solutions for urban navigation challenges. This work underscores the potential of combining real-world and virtual data to advance LiDAR odometry and highlights its role for the future smart cities.
Guided Exploration Reinforcement Learning for 3D UAV Pursuit-evasion Games
No full textWith the increasing use of unmanned aerial vehicles (UAVs) in modern warfare, chasing and colliding with enemy UAVs using a quadrotor has become a promising countermeasure. This highlights the need to solve 3D UAV pursuit-evasion problems. Traditional pursuit algorithms, such as proportional navigation guidance (PNG), struggle to perform effectively in scenarios involving dynamic delays, sensor lags, and moving targets. Therefore, we propose a guided exploration deep deterministic policy gradient (GE-DDPG) to effectively solve the 3D pursuit-evasion problem for quadrotor UAVs. To this end, we propose a guided exploration method to improve training efficiency and performance by leveraging the expert strategy without requiring large demonstration datasets. In addition, a new guidance law based on minimizing the relative distance is proposed for the expert policy of quadrotor UAVs. Through simulation results, the proposed method outperforms the classical guidance law and baseline RL/imitation learning in scenarios involving both stationary and moving evasion UAVs. Moreover, guided exploration remains robust and matches the performance of a shaped dense reward with privileged information even with degraded expert accuracy in non-ideal environments. We also introduce noise scheduling to avoid deadlock situations and learning failures, paving the way for reliable policy transfer. Finally, our trained policy demonstrates robustness against uncertainties in dynamic lag and sensor delay, suggesting its potential for real-world deployment.
Bioelectrosynthesis of Signaling Molecules for Selective Modulation of Cell Signaling
No full textBioelectrosynthesis holds great potential for studying and regulating biological systems through the in situ synthesis and delivery of cell signaling molecules with high spatiotemporal precision. Despite recent advancements, precise control over multiple signaling molecules within a single platform remains challenging. Here, we introduce a bioelectrosynthesis approach capable of selectively producing two types of signaling molecules from a single precursor. This system leverages multi-metal sulfide electrocatalysts inspired by denitrifying enzymes, which generate signaling molecules, nitric oxide (NO), and ammonia (NH3) from nitrite ions. By controlling catalytic active sites, NO or NH3 can be selectively produced under mild electric fields in physiologically relevant conditions. In situ product analyses and first-principles calculations reveal that NO intermediate binding affinity determines product selectivity. These electrocatalysts integrate seamlessly with biological systems, allowing precise, on-demand modulation of NO- or NH3-mediated signaling pathways in human cell lines. By combining electrochemical precision with selective cell control, this strategy may advance the study and regulation of biological systems.
Uniform Block Copolymer Photonic Microspheres with Tunable Structural Colors across Full Visible Spectrum Enabled by Quaternizing Additives
No full textBlock copolymer (BCP) photonic microspheres have attracted significant interest for applications in inks, sensors, and displays. However, their fabrication typically requires bottlebrush BCPs with high molecular weights to attain domain sizes sufficient for visible structural color. In this study, we report a straightforward and effective strategy for producing uniform-sized photonic microspheres through the confined assembly of lamellae-forming poly(styrene-block-2-vinylpyridine) (PS-b-P2VP) diblock copolymers and quaternizing additives within emulsion droplets. This approach allows for tunable structural color across the entire visible spectrum. A series of bromine-functionalized additives are synthesized and employed to increase BCP domain sizes. Among these, 8-bromooctylbenzene (BOB) is identified as the most effective additive due to its exceptional reactivity toward P2VP chains, enabling efficient quaternization of the BCPs. The incorporation of BOB induces a significant increase of domain spacing from 147 to 212 nm while maintaining well-ordered, onion-like lamellae in BCP microspheres. This method enables precise control over the structural colors of BCP photonic microspheres across the whole visible range by simply adjusting the molar ratio of BOB to BCP. Furthermore, the BCP photonic microspheres exhibit reversible structural color transitions in response to pH variations, highlighting their potential for advanced sensing applications.
Femtosecond real-time fragmentation dynamics of the nitrobenzene anion reveal the dissociative electron attachment mechanism
No full textThe femtosecond real-time dynamics of the nitrobenzene anion (C6H5NO2-) in the excited state have been investigated using a recently developed time-resolved photofragment depletion (TRPD) spectroscopic technique, providing molecular-level insight into the C-N bond dissociation pathway leading to (center dot)C6H5 and NO2- fragments for the first time. Ultrafast electronic relaxation from the D2 state, prepared at 2.48 eV, to the ground state (D0) is followed by statistical unimolecular dissociation, yielding NO2- with a lifetime (tau) of approximately 294 ps. This behavior stands in stark contrast to the prompt bond rupture typically observed in conventional dissociative electron attachment (DEA) processes, offering deep insight into the energy flow that governs anionic bond dissociation following electron-molecule collisions.
Advances and Trends in Space Robotics Technology Development and Research
No full textIn recent years, outer space has become a platform for creating new services and value, while also emerging as a competitive arena among nations and various groups. This competition has driven the increasing complexity and advancement of space missions and services, leading to the realization of diverse space service, such as satellite life extension, large-scale solar plant, and space habitat. This paper aims to analyze recent developments in space robotics technology and identify key space robots to provide insights for domestic technological advancement. The study classifies and identifies major space robots deployed or planned for deployment in the past two decades, including lunar, Martian, asteroid exploration, and orbital mission robots, and examines their development outcomes and operational performance. Additionally, the paper provides a guide based on trends among these robots, covering optimal weight, size, and configuration, while summarizing key information on operational conditions and component selection for reference.
대향류 연소 해석을 통한 고체연료의 후퇴율 및 연소 특성 예측에 관한 연구
No full textThis study investigates the regression rate and combustion characteristics of solid fuel using a simple one-dimensional counterflow diffusion flame analysis. A coupled computational framework was developed by integrating gas-phase and condensed-phase reaction models. Simulation results were validated using regression rates from counterflow combustion experiments. Furthermore, the effects of oxygen concentration, ambient pressure, and oxidizer momentum flux on the combustion behaviors of solid fuel were systematically examined. The predicted regression rates showed good agreement with experimental data across the range of conditions tested in this study, although discrepancies were observed at lower pressures. An increase in oxygen concentration, ambient pressure, and oxidizer momentum flux was found to enhance the regression rate. In particular, the increase in momentum flux and pressure shifts the peak temperature location closer to the fuel surface. This proximity intensifies heat feedback to the solid surface, thereby accelerating the pyrolysis process and increasing the regression rate. Furthermore, the correlation analysis identified the oxygen mass fraction as the most influential of the three examined parameters affecting the regression rate. These findings offer fundamental insight into the pyrolysis and combustion of solid fuel relevant to air-breathing solid fuel propulsion systems.
T-CLASS: An Online Tool for the Identification and Classification of Aging and Senescence Using Transcriptome Data
No full textTranscriptome analysis has become increasingly utilized in aging research. However, the identification of the key molecular changes underlying aging processes and longevity-promoting regimens from transcriptome data remains challenging. Here, we present Transcriptomic CLassification via Adaptive learning of Signature States (T-CLASS), an online tool that identifies, from transcriptome data, gene sets of several hundred genes that provide an optimal representation of longevity and aging paradigms. We systematically evaluated the effectiveness of T-CLASS with diverse datasets, including longevity-promoting regimens in Caenorhabditis elegans, cellular senescence by different means in both cultured mouse primary cells and cultured human cells, and human sarcopenia. We found that T-CLASS exhibited robust and high classification performance across datasets compared to preexisting machine/deep learning-based gene selection tools. By focusing our further analysis on longevity-promoting regimens in C. elegans, we showed that T-CLASS successfully classified transcriptomic changes caused by ten lifespan-extending small molecules, among which we experimentally validated the effect of rifampicin and atracurium as a proof of principle. Overall, T-CLASS is an effective and practical tool for uncovering and classifying physiological changes caused by genetic and pharmacological interventions that affect aging.
