Hong Kong University of Science and Technology
Hong Kong University of Science and Technology Institutional RepositoryNot a member yet
162821 research outputs found
Sort by
Improving model fusion by training-time neuron alignment with fixed neuron anchors
No full textModel fusion aims to integrate several deep neural network (DNN) models' knowledge into one by fusing parameters, and it has promising applications, such as improving the generalization of foundation models and parameter averaging in federated learning. However, models under different settings (data, hyperparameter, etc.) have diverse neuron permutations; in other words, from the perspective of loss landscape, they reside in different loss basins, thus hindering model fusion performances. To alleviate this issue, previous studies highlighted the role of permutation invariance and have developed methods to find correct network permutations for neuron alignment after training. Orthogonal to previous attempts, this paper studies training-time neuron alignment, improving model fusion without the need for post-matching. Training-time alignment is cheaper than post-alignment and is applicable in various model fusion scenarios. Starting from fundamental hypotheses and theorems, a simple yet lossless algorithm called TNA-PFN is introduced. TNA-PFN utilizes partially fixed neuron weights as anchors to reduce the potential of training-time permutations, and it is empirically validated in reducing the barriers of linear mode connectivity and multi-model fusion. It is also validated that TNA-PFN can improve the fusion of pretrained models under the setting of model soup (vision transformers) and ColD fusion (pretrained language models). Based on TNA-PFN, two federated learning methods, FedPFN and FedPNU, are proposed, showing the prospects of training-time neuron alignment. FedPFN and FedPNU reach state-of-the-art performances in federated learning under heterogeneous settings and can be compatible with the server-side algorithm.</p
Early diagnosis of nasopharyngeal carcinoma based on machine learning modelling and blood plasma metallomics analysis
No full textCurrent diagnosis of nasopharyngeal carcinoma (NPC) mainly relies on detection of plasma Epstein-Barr virus DNA or nasal endoscopy. However, trace metals or other elements may have critical roles in the pathophysiology of NPC. In this pilot study, blood plasma samples from 93 NPC patients and 30 healthy control were prepared by alkali dilution method, and metal contents were analysed quantitatively by inductively coupled plasma-mass spectrometry. We then built six machine learning (ML) algorithms based on the element concentrations in blood plasma and evaluated the predictive performance by the area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations was employed to interpret the prediction results and explain the contribution of each variable to the model. Compared to the healthy control group, patients with NPC were characterised by increased tin (p < 0.01) and reduced in nickel and iron (p < 0.05), phosphorus (p < 0.01), magnesium, manganese, cobalt, zinc, strontium, molybdenum, antimony, barium, thallium and lead (p < 0.001) concentrations in the plasma. Among the ML models, the bagging model demonstrated the most promising performance in discriminating NPC patients with AUC of 0.999 in testing sets. We further recruited 15 patients with esophageal squamous cell carcinoma (ESCC), 15 non-cancer patients, and used them as blind testing samples. The model can successfully identify NPC patients from those samples with AUC and specificity of 0.857 and 0.800. In summary, the present pilot study highlights the use of metallomics analysis combined with machine learning in NPC identification, especially in early-stage cancer prediction.</p
Miners' Reward Elasticity and Stability of Competing Proof-of-Work Cryptocurrencies
No full textProof-of-Work cryptocurrencies employ miners to sustain the system through algorithmic reward adjustments. We develop a stochastic model of the multicurrency mining and identify conditions for stable transaction speeds. Bitcoin's algorithm requires hash supply elasticity (Formula presented.) 1 for stability, while ASERT remains stable for any elasticity and can be interpreted as a form of stochastic gradient descent. Interactions with other currencies can relax Bitcoin's stability requirements. Using a halving event, we estimate miners' hash supply elasticity and conduct counterfactual simulations. Our findings reveal Bitcoin's heavy reliance on low hash supply elasticity and interactions with smaller cryptocurrencies, urging an algorithm upgrade for stability.</p
Tailoring biochar for volatile organic compounds adsorption via molten salt–assisted biomass pyrolysis
No full textMolten salt–assisted pyrolytic carbonization was applied to convert biomass into functional biochar for volatile organic compound (VOC) adsorption. By varying molten salt types and salt-to-feedstock ratios, biochar with distinct pore structures and surface compositions was obtained. The adsorption behaviors of benzene and toluene were evaluated at 283, 298 and 313 K. For both VOCs, adsorption uptakes increased with partial pressure, while benzene adsorption decreased with increasing temperature, indicating VOC uptake is primarily governed by pore-controlled physical adsorption. Carbonate-assisted biochar (PSL-1-800 and PSL-3-800) exhibited higher adsorption capacities than other samples, which were closely associated with their enhanced microporosity and surface area. Density functional theory (DFT) calculations and wavefunction analyses indicate that adsorption on metal-free biochar is governed by weak π–π interactions, whereas metal-containing biochar exhibit strengthened adsorption through cation–π interactions. Together with the experimental observations, these results show that molten salt chemistry controls VOC adsorption primarily by regulating microporous structure and metal-assisted surface interactions. This study demonstrates molten salt–assisted pyrolysis of biomass waste is a promising approach to develop efficient VOC adsorbents, offering a promising proof-of-concept strategy for developing biochar-based adsorbents for VOC removal.</p
FedCHG: Graph autoencoder enhanced federated learning for cross-Domain heterogeneous graph
No full textFederated Graph Learning is an efficient technique for processing graph data, capable of integrating information from various data sources in a distributed environment. However, graph data from different domains often exhibit significant differences in both features and structures, which is known as the heterogeneity problem of cross-domain federated graph data. Existing methods still have limitations in the underlying extraction and aggregation of structural information for cross-domain federated graphs. To address this issue, we propose FedCHG: Graph Autoencoder Enhanced Federated Learning for Cross-Domain Heterogeneous Graph. Specifically, we combine the Graphlet subgraph algorithm and the random walk diffusion algorithm to extract graph structural information from both local and global perspectives, constructing a universal structural representation. We design a graph encoder-decoder architecture to integrate global structural information while preserving local structural information. Additionally, we introduce a weighted aggregation strategy based on graph structural similarity, which considers structural differences during the aggregation process and enhances the overall performance of the model. Experiments on publicly available multi-domain graph datasets show that compared to current state-of-the-art federated graph learning baselines, the proposed FedCHG improves the accuracy by up to 5% in cross-domain scenarios, demonstrating the effectiveness of FedCHG in collaborative learning scenarios in various graph datasets.</p
Hybrid resolved-unresolved CFD-DEM framework for multiscale fluid-particle systems with irregular-shaped and polydisperse particles
No full textThis study presents a hybrid resolved and unresolved computational fluid dynamics-discrete element method (CFD-DEM) coupling framework for modeling fluid-particle systems involving irregularly shaped and polydisperse particles. The framework integrates high-fidelity signed distance field (SDF) representations for coarse particles with efficient coarse-fine contact algorithms and unresolved or semi-resolved fluid-particle interaction schemes for fine particles. Key features such as accurate representation of irregularly shaped particles, consideration of varying particle sizes, and evaluation of fluid-particle interactions in complex granular settings are highlighted. The hybrid CFD-DEM solver is implemented and verified through a series of standard benchmark tests. A detailed filtration example is provided to demonstrate its capability in modeling multiscale fluid-particle interactions and in analyzing the influence of particle size and shape on filtration behavior. Additional case studies, including channelized sorting, jet-induced destabilization, and dam-break flows, further illustrate the flexibility and effectiveness of the proposed approach in practical applications.</p
Downside risk similarity and M&As
No full textDownside risks are ubiquitous and can profoundly impact firm operations and valuation. Failure to adequately assess and manage target firms' downside risks hinders acquirers' ability to integrate and manage these businesses. This article introduces a novel measure of firms' downside risk similarity (DRS) based on risk factor descriptions and examines its implications for mergers and acquisitions (M&A) outcomes. We first validate that the measure is distinct from existing similarity measures and that it captures similarity in firms' potential significant downside. Using the new measure, we find that the market reacts more positively to deals in which acquirers and targets share more downside risks. Additional analyses show that this beneficial effect of DRS is driven primarily by risks that are idiosyncratic or firm-specific, consistent with these risks requiring acquirers' relevant expertise to manage. Last, we document that in deals with more similar downside risks, the acquirers experience fewer risk profile changes and are less likely to suffer from adverse outcomes, such as deal-specific goodwill impairment, divestitures, and significant profitability declines. Overall, we conclude that DRS plays a significant role in the M&A process.</p
Full-Color and Switchable Phosphorescence of Carbene-Metal-Amide-Based Bimetallic Gold(I) Complexes with Dynamic through-Space Interaction
No full textRegulating through-space interactions offers a promising strategy for designing multifunctional luminescent materials. However, integrating stimuli-responsive photophysical behaviors into such systems remains challenging. In this study, a series of carbene-metal-amide bimetallic Au(I) complexes featuring dynamic through-space interactions is reported that enable aggregation-induced emission and full-color-tunable photoactivated phosphorescence. Single-crystal X-ray diffraction combined with theoretical calculations reveals conformationally adaptive frameworks that facilitate ligand rotational freedom (carbazole) and N-heterocyclic carbene conformational flexibility, enabling precise modulation of intramolecular through-space interactions. These complexes exhibit multi-stimuli-responsive phosphorescence, allowing reversible, on-demand switching of emission color and intensity across molecular and macroscopic scales. By strategically blending phosphors, white-light emission with a CIE 1931 coordinate of (0.30, 0.31) is achieved. The materials further demonstrate time-resolved information encryption capabilities, making them ideal for light-activated printing and high-security anti-counterfeiting inks. This work advances the rational design of smart luminescent platforms for applications in optoelectronics, sensing, and photonic security.</p
Fractional-order stochastic resonance-based rescaling-frequency scanning images for early multi-frequency fault detection of machines
No full textIn engineering applications, weak multi-frequency fault signals from mechanical equipment are often masked by strong background noise. Traditional stochastic resonance (SR) methods mainly focus on enhancing fault signals into sine-like ones, but they may lose or even destroy the multi-harmonic characteristics of fault signals. To this end, this paper would propose a rescaling-frequency scanning image method using fractional-order SR (FSR-RFSI), aiming to enhance and visualize weak multi-frequency useful signals. First, the proposed method develops a fractional-order SR system with memory properties, which is designed to detect weak multi-frequency signals in complex spectral environments. Moreover, a weighted zero-crossing signal-to-noise ratio (WZCSNR) is proposed as a performance evaluation metric, which effectively overcomes the limitation of the traditional signal-to-noise ratio (SNR) that focuses solely on frequency-domain energy while neglecting time-domain multi-harmonic components. Meanwhile, to improve parameter tuning efficiency, this paper establishes an analytical relationship map between the resonant frequency and system parameters, namely rescaling-frequency scanning image. Furthermore, a quantum genetic algorithm (QGA) is used to achieve adaptive optimization of key system parameters. Simulation analyses and experiments on early rolling bearing and gearbox faults show that the proposed method can effectively boost and detect weak multi-frequency fault signals. Additionally, comparative analysis with Maximum Correlated Kurtosis Deconvolution (MCKD), Fast Kurtogram (FK), and Feature Modal Decomposition (FMD) methods further validates the superiority of the proposed method.</p