The proposition of lottery ticket hypothesis revealed the relationship between network structure and initialization parameters and the learning potential of neural networks. The original lottery ticket hypothesis performs pruning and weight resetting
after training convergence, exposing it to the problem of forgotten learning knowledge and potential high cost of training. Therefore, we propose a strategy that combines the idea of neural network structure search with a pruning algorithm to alleviate this problem. This algorithm searches and extends the network structure on existing winning ticket sub-network to producing new winning ticket recursively. This allows the training and pruning process to continue without compromising performance. A new winning ticket sub-network with deeper network structure, better generalization ability and better test performance can be obtained in this recursive manner. This method can solve: the difficulty of training or performance degradation of the sub-networks after pruning, the forgetting of the weights of the original lottery ticket hypothesis and the difficulty of generating winning ticket sub-network when the final network structure is not given. We validate this strategy on the MNIST and CIFAR-10 datasets. And after relating it to similar biological phenomena and relevant lottery ticket hypothesis studies in recent years, we will further propose a new hypothesis to discuss which factors that can keep a network juvenile, i.e., those possible factors that influence the learning potential or generalization performance of a neural network during training.
Out-of-distribution (OOD) detection and lossless compression constitute two problems that can be solved by the training of probabilistic models on a first dataset with subsequent likelihood evaluation on a second dataset, where data distributions dif
fer. By defining the generalization of probabilistic models in terms of likelihood we show that, in the case of image models, the OOD generalization ability is dominated by local features. This motivates our proposal of a Local Autoregressive model that exclusively models local image features towards improving OOD performance. We apply the proposed model to OOD detection tasks and achieve state-of-the-art unsupervised OOD detection performance without the introduction of additional data. Additionally, we employ our model to build a new lossless image compressor: NeLLoC (Neural Local Lossless Compressor) and report state-of-the-art compression rates and model size.
In indoor positioning, signal fluctuation is highly location-dependent. However, signal uncertainty is one critical yet commonly overlooked dimension of the radio signal to be fingerprinted. This paper reviews the commonly used Gaussian Processes (GP
) for probabilistic positioning and points out the pitfall of using GP to model signal fingerprint uncertainty. This paper also proposes Deep Gaussian Processes (DGP) as a more informative alternative to address the issue. How DGP better measures uncertainty in signal fingerprinting is evaluated via simulated and realistically collected datasets.
Tailoring electron transfer dynamics across solid-liquid interfaces is fundamental to the interconversion of electrical and chemical energy. Stacking atomically thin layers with a very small azimuthal misorientation to produce moire superlattices ena
bles the controlled engineering of electronic band structures and the formation of extremely flat electronic bands. Here, we report a strong twist angle dependence of heterogeneous charge transfer kinetics at twisted bilayer graphene electrodes with the greatest enhancement observed near the magic angle (~1.1 degrees). This effect is driven by the angle-dependent tuning of moire-derived flat bands that modulate electron transfer processes with the solution-phase redox couple. Combined experimental and computational analysis reveals that the variation in electrochemical activity with moire angle is controlled by atomic reconstruction of the moire superlattice at twist angles <2 degrees, and topological defect AA stacking regions produce a large anomalous local electrochemical enhancement that cannot be accounted for by the elevated local density of states alone. Our results introduce moire flat band materials as a distinctively tunable paradigm for mediating electrochemical transformations.
The IceCube Neutrino Observatory at the South Pole detects Cherenkov light emitted by charged secondary particles created by primary neutrino interactions. Double pulse waveforms can arise from charged current interactions of astrophysical tau neutri
nos with nucleons in the ice and the subsequent decay of tau leptons. The previous 8-year tau double pulse analysis found three tau neutrino candidate events. Among them, the most promising one observed in 2014 is located very near the dust layer in the middle of the detector. A posterior analysis on this event will be presented in this paper, using a new ice model treatment with continuously varying nuisance parameters to do the targeted Monte Carlo re-simulation for tau and other background neutrino ensembles. The impact of different ice models on the expected signal and background statistics will also be discussed.
In recent years, many deep learning models have been adopted in autonomous driving. At the same time, these models introduce new vulnerabilities that may compromise the safety of autonomous vehicles. Specifically, recent studies have demonstrated tha
t adversarial attacks can cause a significant decline in detection precision of deep learning-based 3D object detection models. Although driving safety is the ultimate concern for autonomous driving, there is no comprehensive study on the linkage between the performance of deep learning models and the driving safety of autonomous vehicles under adversarial attacks. In this paper, we investigate the impact of two primary types of adversarial attacks, perturbation attacks and patch attacks, on the driving safety of vision-based autonomous vehicles rather than the detection precision of deep learning models. In particular, we consider two state-of-the-art models in vision-based 3D object detection, Stereo R-CNN and DSGN. To evaluate driving safety, we propose an end-to-end evaluation framework with a set of driving safety performance metrics. By analyzing the results of our extensive evaluation experiments, we find that (1) the attacks impact on the driving safety of autonomous vehicles and the attacks impact on the precision of 3D object detectors are decoupled, and (2) the DSGN model demonstrates stronger robustness to adversarial attacks than the Stereo R-CNN model. In addition, we further investigate the causes behind the two findings with an ablation study. The findings of this paper provide a new perspective to evaluate adversarial attacks and guide the selection of deep learning models in autonomous driving.
In this paper, we accomplish the complete one-loop matching of the type-I seesaw model onto the Standard Model Effective Field Theory (SMEFT), by integrating out three heavy Majorana neutrinos with the functional approach. It turns out that only 31 d
imension-six operators (barring flavor structures and Hermitian conjugates) in the Warsaw basis of the SMEFT can be obtained, and most of them appear at the one-loop level. The Wilson coefficients of these 31 dimension-six operators are computed up to $mathcal{O}left( M^{-2}right)$ with $M$ being the mass scale of heavy Majorana neutrinos. As the effects of heavy Majorana neutrinos are encoded in the Wilson coefficients of these higher-dimensional operators, a complete one-loop matching is useful to explore the low-energy phenomenological consequences of the type-I seesaw model. In addition, the threshold corrections to the couplings in the Standard Model and to the coefficient of the dimension-five operator are also discussed.
In this paper, we propose and analyze a diffuse interface model for inductionless magnetohydrodynamic fluids. The model couples a convective Cahn-Hilliard equation for the evolution of the interface, the Navier-Stokes system for fluid flow and the po
ssion quation for electrostatics. The model is derived from Onsagers variational principle and conservation laws systematically. We perform formally matched asymptotic expansions and develop several sharp interface models in the limit when the interfacial thickness tends to zero. It is shown that the sharp interface limit of the models are the standard incompressible inductionless magnetohydrodynamic equations coupled with several different interface conditions for different choice of the mobilities. Numerical results verify the convergence of the diffuse interface model with different mobilitiess.
Fifth generation (5G) aims to connect massive devices with even higher reliability, lower latency and even faster transmission speed, which are vital for implementing the e-health systems. However, the current efforts on 5G e-health systems are still
not enough to accomplish its full blueprint. In this article, we first discuss the related technologies from physical layer, upper layer and cross layer perspectives on designing the 5G e-health systems. We afterwards elaborate two use cases according to our implementations, i.e., 5G e-health systems for remote health and 5G e-health systems for Covid-19 pandemic containment. We finally envision the future research trends and challenges of 5G e-health systems.
This paper investigates the energy efficiency (EE) optimization in downlink multi-cell massive multiple-input multiple-output (MIMO). In our research, the statistical channel state information (CSI) is exploited to reduce the signaling overhead. To m
aximize the minimum EE among the neighbouring cells, we design the transmit covariance matrices for each base station (BS). Specifically, optimization schemes for this max-min EE problem are developed, in the centralized and distributed ways, respectively. To obtain the transmit covariance matrices, we first find out the closed-form optimal transmit eigenmatrices for the BS in each cell, and convert the original transmit covariance matrices designing problem into a power allocation one. Then, to lower the computational complexity, we utilize an asymptotic approximation expression for the problem objective. Moreover, for the power allocation design, we adopt the minorization maximization method to address the non-convexity of the ergodic rate, and use Dinkelbachs transform to convert the max-min fractional problem into a series of convex optimization subproblems. To tackle the transformed subproblems, we propose a centralized iterative water-filling scheme. For reducing the backhaul burden, we further develop a distributed algorithm for the power allocation problem, which requires limited inter-cell information sharing. Finally, the performance of the proposed algorithms are demonstrated by extensive numerical results.
