Search

Deep neural networks complete a feature extraction task by propagating the inputs through multiple modules. However, how the representations evolve with the gradient-based optimization remains unknown. Here we leverage the intrinsic dimension of the representations to study the learning dynamics and find that the training process undergoes a phase transition from expansion to compression under disparate training regimes. Surprisingly, this phenomenon is ubiquitous across a wide variety of model architectures, optimizers, and data sets. We demonstrate that the variation in the intrinsic dimension is consistent with the complexity of the learned hypothesis, which can be quantitatively assessed by the critical sample ratio that is rooted in adversarial robustness. Meanwhile, we mathematically show that this phenomenon can be analyzed in terms of the mutable correlation between neurons. Although the evoked activities obey a power-law decaying rule in biological circuits, we identify that the power-law exponent of the representations in deep neural networks predicted adversarial robustness well only at the end of the training but not during the training process. These results together suggest that deep neural networks are prone to producing robust representations by adaptively eliminating or retaining redundancies. The code is publicly available at https://github.com/cltan023/learning2022. [ABSTRACT FROM AUTHOR]

The advancement of integrated optical communication networks necessitates the deployment of on-chip beam splitters for efficient signal interconnections at network nodes. However, the pursuit of micron-scale beam splitting with large corners and reducing the device footprint to boost connection flexibility often results in phase mismatches. These mismatches, which stem from radiation modes and backward scattering, pose significant obstacles in creating highly integrated and interference-resistant connections. To address this, we introduce a solution based on the topological valley-contrasting state generated by photonic crystals with opposing valley Chern numbers, manifested in a harpoon-shaped structure designed to steer the splitting channels. This approach enables adiabatic mode field evolution over large corners, capitalizing on the robust phase modulation capabilities and topological protection provided by the subwavelength-scale valley-contrasting state. Our demonstration reveals that beam splitters with large corners of 60°, 90°, and 120° exhibit insertion loss fluctuations below 2.7 dB while maintaining a minimal footprint of 8.8 µm × 8.8 µm. As a practical demonstration, these devices facilitate three-channel signal connections, successfully transmitting quadrature phase shift keying signals at 3.66 Tbit/s with bit error rates below the forward error correction threshold, demonstrating performance comparable to that in defects scenarios. By harnessing the unidirectional excitation feature, we anticipate significant enhancements in the capabilities of signal distribution and connection networks through a daisy chain configuration. [ABSTRACT FROM AUTHOR]

In this study, we explored the neural mechanism of global topological perception in the human visual system. We showed strong evidence that the retinotectal pathway in the archicortex of the human brain is responsible for global topological perception, and for modulating the local feature processing in the classical ventral visual pathway. Inspired by this recent cognitive discovery, we developed a novel CogNet architecture to emulate the global-local dichotomy of human visual cognitive mechanisms. The thorough experimental results indicate that the proposed CogNet not only significantly improves image classification accuracies but also effectively addresses the texture bias problem observed in baseline CNN models. We have also conducted mathematical analysis for the generalization gap for general neural networks. Our theoretical derivations suggest that the Hurst parameter, a measure of the curvature of the loss landscape, can closely bind the generalization gap. A larger Hurst parameter corresponds to a better generalization ability. We found that our proposed CogNet achieves a lower test error and attains a larger Hurst parameter, strengthening its superiority over the baseline CNN models further. [ABSTRACT FROM AUTHOR]

Clustering high-dimensional data is very challenging due to the curse of dimensionality. To address this problem, low-rank matrix approximations are widely used to identify the underlying low-dimensional structure of a dataset. Among these, nonnegative matrix factorization (NMF) is the most popular because its decomposed factors are nonnegative and meaningful. However, the NMF problem has been proved to be nonconvex and NP-hard, thus resulting in many local minima. To obtain high-quality local minima, we propose an entropy regularized fuzzy nonnegative matrix factorization (ERF-NMF) model for high-dimensional data fuzzy clustering. First, probability simplex constraints on the decomposed weight components are added to achieve dimension reduction and fuzzy clustering of a dataset simultaneously. Based on the constraints, we also introduce entropy regularization to further reduce the search space for optimal solutions. Finally, we present multiplicative update rules for solving the ERF-NMF model and provide a complexity and convergence analysis. Comprehensive experiments show that the proposed ERF-NMF performs remarkably well with promising results, and its decomposition will be sparser because of entropy regularization and have a clearer physical meaning because of probability simplex constraints. [ABSTRACT FROM AUTHOR]

In this paper, a fuzzy dual-hunting control based on auction algorithm is proposed. Relying on a single containment gives the small target the possibility of escape. Therefore, we design a dual-hunting framework for multiagent systems to hunt a small target. A hunting formation with double containment is designed to reduce the possibility of the target escaping. Then, an auction algorithm is designed to generate the desired formation and rationally plan the positions of the agents in the hunting formation. To improve the applicability of the dual-hunting framework, we take high-order multiagent systems as the research object, and the fuzzy logic system (FLS) is introduced to approximate the unknown nonlinear dynamics (UND) due to higher-order system model errors and environmental disturbances. Based on FLS and backstepping method, a hunting controller for high-order systems is designed, which enables multiagent systems to track targets and form hunting formations. Furthermore, the Lyapunov function is designed to prove the stability of the controller. Finally, the effectiveness of the proposed method is demonstrated by simulating a multiagent system containing 12 intelligences hunting for a dynamic target. [ABSTRACT FROM AUTHOR]

In this paper, the optimal incremental-containment control of two-order swarm system based on reinforcement learning (RL) is proposed to avoid the dilemma that the number of agents in a swarm system is immutable, which is essential for a swarm system that cannot meet the containment demands and need more agents to expand the containment range. Notably, the number of agents in a swarm system with a traditional containment controller is immutable, which limits the containment range that the swarm system can achieve. Besides, in traditional optimal control theory, it is obtained by solving the Hamilton-Jacobi-Bellman (HJB) equation, which is difficult to solve due to the unknown nonlinearity. To overcome these problems, several contributions are made in this paper. Firstly, in order to overcome the dilemma that the number of agents in the swarm system is immutable, the incremental-containment control is proposed. Secondly, considering the error and control input as the optimization goal, the optimal cost function is introduced and the optimal incremental-containment control is proposed to reduce resource waste and increase hardware service life. Furthermore, based on the proposed optimal incremental-containment control, the controller is designed by a new RL based on the backstepping method. The Lyapunov function is used to prove the stability of controller. The simulation show the efficiency of the proposed controller. [ABSTRACT FROM AUTHOR]

There is a growing application of machine learning methods to predict socioeconomic and environmental attributes in computational social science, where big data are usually presented in tabular format. However, it is still a challenge to develop novel deep learning models to deal with tabular data, fill missing value, improve prediction accuracy, and enhance interpretability. In this study, we for the first time apply a tabular deep learning methodology (TabNet) to predict socioeconomic and environmental attributes (number of population and companies, volume of consumption, poker players' behaviors, forest cover, etc.). Furthermore, we develop a new network architecture, referred to as improved TabNet (iTabNet), that can simultaneously learn local and global features in the tabular data to improve prediction accuracy. We also introduce a difference loss to constrain the feature selection process in iTabNet so that the model can use different features at different steps to enhance interpretability. To deal with missing values, we introduce a fusion strategy based on data mean and Auto-Encoder network to efficiently complete a more reasonable value filling. Experimental results demonstrate that the proposed iTabNet achieves competitive performances in the application to predict socioeconomic and environmental attributes based on tabular data, iTabNet using the proposed fusion strategy significantly outperforms other machine learning models when tabular data have missing values. [ABSTRACT FROM AUTHOR]

This paper develops a groupwise dimension reduction-based adaptive-to-model test for partially linear single-index models. The test behaves as a local smoothing test would if the model were bivariate. The test statistic under the null hypothesis is asymptotically normally distributed. The test can detect local alternatives distinct from the null hypothesis at the rate that existing local smoothing tests can achieve when the regression model contains bivariate covariates. Therefore, the curse of dimensionality is largely alleviated. Numerical studies, including two real data examples, are conducted to examine the finite sample performance of the proposed test. [ABSTRACT FROM AUTHOR]

Purpose: To address whether Curcumin has synergistic effect with cytarabine (Ara-C) in treating acute myeloid leukemia (AML). Methods: A xenograft AML mouse model was established by injecting HL-60 cells into tail vein of mice to assess the function of Curcumin. Mononuclear cells (MNCs) isolated from AML mice and AML cell lines were used to examine the effect of Curcumin. Metagenomics and metabolomics were used to evaluate the alteration of intestinal microbiota and the change of metabolites in MNCs. Results: Curcumin treatment sensitized response to Ara-C in MNCs of AML mice, but had no direct effect on AML cell lines. Metagenomics revealed an alteration of intestinal microbiota with Curcumin treatment, which contributes to sensitized response to Ara-C. Curcumin treatment led to enhanced intestinal intact to sensitize response to Ara-C in AML mice, through reducing mucus degrading bacteria. Metabolomics demonstrated that Curcumin treatment led to decreased cholesterol in MNCs of AML mice. Further study proved that Curcumin treatment resulted in inhibition of SQLE, a key enzyme of cholesterol biosynthesis, to increase sensitivity to Ara-C. Conclusion: Curcumin sensitizes response to Ara-C through regulating microbiota, highlighting the importance of intestinal intact strengthening in chemoresistant therapy. Moreover, aiming at cholesterol synthesis is promising in AML treatment. [ABSTRACT FROM AUTHOR]

Speech synthesis, an artificial intelligence technology that employs computers to imitate human speech, has played a crucial role in human–computer interaction since it can automatically convert text into speech with satisfactory intelligibility and naturalness. Tacotron2 is the second generation end-to-end English speech synthesis model developed by Google. As Mandarin becomes more and more popular in the world, the associated speech synthesis technologies have been applied in various applications. Aiming at extending Tacotron2 to synthesize Mandarin speech, we propose in this paper a novel synthesis method by adding a Mandarin-to-PinYin module and a prosodic structure prediction model into Tacotron2. By evaluating synthesized results with subjective and objective methods, the added prosodic structure prediction model is demonstrated to help Tacotron2 synthesize more natural and human-like Mandarin speech. [ABSTRACT FROM AUTHOR]

Six organophosphate esters (OPEs) with different octanol-air partition coefficient (KOA) were measured from gas and particle samples collected from an urban area of Central China between June 2018 and May 2019. The highest concentration of total OPEs (∑6OPEs) in total suspended particle (TSP) were found in winter, followed by spring, autumn, and summer. ∑6OPEs showed significant positive correlation with TSP concentration (p < 0.01), significant negative correlation with temperature (p < 0.01), weak negative linear correlations with relative humility (p < 0.05), and no significant correlations with wind speed (p > 0.05). Although OPEs were not detected in polyurethane foams (PUFs), the poly-parameter linear free energy relationship model (pp-LFER) estimated particle fractions of OPEs (fpart pp-LFERg/p) were below 2.73% for Tris (2-chloroisopropyl) phosphate (TCIPP), Tris (chloroethyl) phosphate (TCEP), and Tributyl phosphate (TnBP); 67.8% for Triphenyl phosphate (TPhP); and above 99.6% for TBOEP and TCrP. The concentrations of particle-bound TPhP, TCrP, and TBOEP estimated by pp-LFER (Cpart pp-LFERg/p) were 0.19, 0.09, and 0.05 ng/m3, which were in good agreement with measurements. However, the estimated Cpart pp-LFERg/p of TCEP, TCIPP, and TnBP showed obvious differences with measured data. According to filter-air partitioning estimated by pp-LFER, the strong adsorption of gaseous OPEs to glass fiber filter (GFF) may be responsible to the underestimation. Besides, the weak adsorption of gaseous OPEs to PUF is likely to be another important reason for the discrepancy between measured and modeled results. Therefore, the risk assessment of gaseous OPEs should be of concern in the further research. [ABSTRACT FROM AUTHOR]

Seismic monitoring is an important verification technique under the Comprehensive Nuclear-Test-Ban Treaty. Phase association technology, which is an important component of seismic data processing, associates signals generated from the same event source recorded at multiple stations and determines event information based on signal features. Seismic event association based on the historical seismic data feature model is a research hot spot in the field of seismic monitoring. In this paper, an event association method called NET-VISA based on Bayesian theory is introduced; then, the application of the historical data feature model in NET-VISA is analyzed. The NET-VISA method is evaluated using the International Data Centre LEB bulletins published by the Comprehensive Nuclear-Test-Ban Treaty Organization, the ISC Reviewed Bulletins, and the China Earthquake Networks Center bulletin as reference sets. The results show that for the global sparse network, NET-VISA is generally superior to the GA method currently used by the IDC, which verifies NET-VISA's effectiveness. However, NET-VISA misses some events detected by the GA. The reasons might be that these events are located in regions with low seismic activity and that insufficient historical event data exists, resulting in unreasonable scoring results.Finally, the application method and research direction of NET-VISA in actual scenarios are discussed. [ABSTRACT FROM AUTHOR]

Background: Recently, sparse unmixing has been successfully applied to spectral mixture analysis of remotely sensed hyperspectral images. Based on the assumption that the observed image signatures can be expressed in the form of linear combinations of a number of pure spectral signatures known in advance, unmixing of each mixed pixel in the scene is to find an optimal subset of signatures in a very large spectral library, which is cast into the framework of sparse regression. However, traditional sparse regression models, such as collaborative sparse regression, ignore the intrinsic geometric structure in the hyperspectral data. Results: In this paper, we propose a novel model, called manifold regularized collaborative sparse regression, by introducing a manifold regularization to the collaborative sparse regression model. The manifold regularization utilizes a graph Laplacian to incorporate the locally geometrical structure of the hyperspectral data. An algorithm based on alternating direction method of multipliers has been developed for the manifold regularized collaborative sparse regression model. Conclusions: Experimental results on both the simulated and real hyperspectral data sets have demonstrated the effectiveness of our proposed model. [ABSTRACT FROM AUTHOR]

Recently, the design of group sparse regularization has drawn much attention in group sparse signal recovery problem. Two of the most popular group sparsity-inducing regularization models are $$\ell _{1,2}$$ and $$\ell _{1,\infty }$$ regularization. Nevertheless, they do not promote the intra-group sparsity. For example, Huang and Zhang (Ann Stat 38:1978-2004, 2010) claimed that the $$\ell _{1,2}$$ regularization is superior to the $$\ell _1$$ regularization only for strongly group sparse signals. This means the sparsity of intra-group is useless for $$\ell _{1,2}$$ regularization. Our experiments show that recovering signals with intra-group sparse needs more measurements than those without, by the $$\ell _{1,\infty }$$ regularization. In this paper, we propose a novel group sparsity-inducing regularization defined as a mixture of the $$\ell _{1/2}$$ norm and the $$\ell _{1}$$ norm, referred to as $$\ell _{1/2,1}$$ regularization, which can overcome these shortcomings of $$\ell _{1,2}$$ and $$\ell _{1,\infty }$$ regularization. We define a new null space property for $$\ell _{1/2,1}$$ regularization and apply it to establish a recoverability theory for both intra-group and inter-group sparse signals. In addition, we introduce an iteratively reweighted algorithm to solve this model and analyze its convergence. Comprehensive experiments on simulated data show that the proposed $$\ell _{1/2,1}$$ regularization is superior to $$\ell _{1,2}$$ and $$\ell _{1,\infty }$$ regularization. [ABSTRACT FROM AUTHOR]

Direct synthesis of ethanol from syngas derived from coal, natural gas, or biomass is one of the most promising routes for renewable energy production. In this work, Ce-promoted highly-dispersed Rh catalyst on TiO 2 support was prepared by the deposition–precipitation method, which exhibits largely enhanced catalytic activity and selectivity for ethanol production from syngas. XRD and TEM reveal a high dispersion of Rh nanoparticles (~1 nm) on the surface of TiO 2 support by the introduction of CeO 2. It was found that the relative content of Rh and CeO 2 imposes significant influence on the catalytic behavior, and the sample of 1 wt% Rh–1 wt% CeO 2/TiO 2 exhibits the highest CO conversion (32 %) and the largest ethanol selectivity (33 %) under the conditions of 573 K, 3 MPa, H 2:CO = 2:1. H 2-TPR and XPS results confirm a gradually enhanced electron interaction between Rh and CeO 2 with the increase of Ce content; CO-TPD further reveals that the Rh–CeO 2 interaction generates a new adsorption site for CO, accounting for the improved selectivity towards ethanol. Therefore, this work provides a facile and effective method for the preparation of Rh–CeO 2/TiO 2 catalysts with high catalytic behavior, which can be potentially used in syngas conversion to ethanol. Graphical Abstract: A series of CeO 2 promoted Rh/TiO 2 catalysts with low Rh loading (1 wt%) were prepared by the deposition–precipitation method, which exhibit largely enhanced catalytic activity and selectivity towards ethanol production from syngas [Figure not available: see fulltext.]. [ABSTRACT FROM AUTHOR]

A novel calix[4]arene-based fluorescent chemosensor bearing a 2-aminopyridine moiety and a naphthalenic fluorophore was synthesized The chemical structure of the product was elucidated by FT-IR, MS-FAB, NMR and elemental analyses. Then, the properties and identification mechanism of the synthesized chemosensor were investigated. The results show that the chemosensor exhibits selective fluorescent quenching in the presence of aromatic organic acid in acetonitrile solution, and that the binding ability of the chemosensor with organic acid is in the order of p-cyanic-benzyl acid > p-chloric-benzyl acid > p-methoxyl-benzyl acid > benzyl acid. [ABSTRACT FROM AUTHOR]

Two composites LaNi4.8Sn0.2/CNTs and NdNi4.8Sn0.2/CNTs were prepared by an impregnation-reduction method. Their hydrogen storage capacity could reach up to 2.96 wt% and 2.88 wt% respectively at room temperature and 1.0 MPa pressure. These values, which might result from the synergetic effect between the alloy nanoparticles and the pretreated CNTs, were three times higher than those of the unsupported MNi4.8Sn0.2(M=La, Nd) alloys under the same conditions. XRD and TEM revealed that the alloy particles were uniformly dispersed on the CNTs and the average particle size was Ca. 30 nm. The composites also showed good stability and the hydrogen storage capacity decreased by less than 6% after 100 adsorption-desorption cycles. Moreover, no noticeable change in crystalline structure was observed for the composites. [ABSTRACT FROM AUTHOR]

A type of novel material with a high hydrogen storage capacity was prepared by supporting PdNi18 alloy nanoparticles, which were synthesized by using a new colloid method, on the surface of pretreated multiwal led carbon nanotubes (MWCNTs). The average PdNi18 alloy particle sizes calculated from XRD patterns were ca. 3 nm, and the high dispersion of these particles on MWCNTs was confirmed by TEM image. Hydrogen storage performance of the composite was investigated under moderate pressure (0.1-1.5 MPa) at room temperature, and a maximum storage capacity of ca. 2.3 wt% was achieved under 1.5 MPa at room temperature, which was much higher than that reported previously under the same conditions. [ABSTRACT FROM AUTHOR]

It is demonstrated by various geochemical indexes that the Zhengjia-Wangzhuang Oilfield with viscous crude oil in the Jiyang Depression has been sourced from the contribution of matured source rocks in the upper Es4. The principal cause leading to the densification of crude oils would be biodegradation, with the degradation level of crude oils being ranked as 2–8; vertically, the biodegradation level increases from the top to bottom of the oil column, with a distinctive biodegradation gradient occurring. Calculated parameters of sterane, terpane and methyl-phenanthrene have indicated that the source-rock's maturity of crude oils and asphaltic sands ranges from 0.7 to 0.9, and based on the calculation of Easy Ro model, the temperature of hydrocarbon generation in the source rock would be within 120– 140°C, which coincides with the measurements of reservoir inclusions. The measured homogenization temperature would represent the generation temperature of the source rock, and be fairly different from that of reservoir while being charged with hydrocarbons, which reflects the hydrocarbon-charging to be a fast process, and the period of pool-formation to be consistent with that of peaked generation. According to the evaluation of generation history, the pool-formation could have been occurring in 7–15 Ma. And the biodegradation of crude oils in the study area would be considered to take place in 4–15 Ma based on the examination of biodegradation order and dynamic calculations. [ABSTRACT FROM AUTHOR]

Presents a letter to the editor on resonances of the winged helix domain of the proto-oncoprotein cQin.