arXiv subjects

Longfei Li

Publications and source records attributed to Longfei Li.

62 records · Page 4Linked to original sources

arXiv2020

Knowledge Consistency between Neural Networks and Beyond

This paper aims to analyze knowledge consistency between pre-trained deep neural networks. We propose a generic definition for knowledge consistency between neural networks at different fuzziness levels. A task-agnostic method is designed to disentangle feature components, which represent the consistent knowledge, from raw intermediate-layer features of each neural network. As a generic tool, our method can be broadly used for different applications. In preliminary experiments, we have used knowledge consistency as a tool to diagnose representations of neural networks. Knowledge consistency provides new insights to explain the success of existing deep-learning techniques, such as knowledge distillation and network compression. More crucially, knowledge consistency can also be used to refine pre-trained networks and boost performance.

cs.LG↗

arXiv2019

Non-Hermitian two-dimensional photonic crystal flat lens

In this paper, a non-Hermitian two-dimensional photonic crystal flat lens is proposed. The negative refraction of the second band of photonic crystal is utilized to realize super-resolution imaging of the point source. Based on the principles of non-Hermitian systems, a negative imaginary part is introduced into the imaging frequency, in which case the imaging intensity and resolution are improved. The results indicate that the non-Hermitian system provides a new method to improve the imaging performance of the photonic crystal lens.

physics.optics↗

arXiv2018

GeniePath: Graph Neural Networks with Adaptive Receptive Paths

We present, GeniePath, a scalable approach for learning adaptive receptive fields of neural networks defined on permutation invariant graph data. In GeniePath, we propose an adaptive path layer consists of two complementary functions designed for breadth and depth exploration respectively, where the former learns the importance of different sized neighborhoods, while the latter extracts and filters signals aggregated from neighbors of different hops away. Our method works in both transductive and inductive settings, and extensive experiments compared with competitive methods show that our approaches yield state-of-the-art results on large graphs.

cs.LG↗

arXiv2018

Distributed Collaborative Hashing and Its Applications in Ant Financial

Collaborative filtering, especially latent factor model, has been popularly used in personalized recommendation. Latent factor model aims to learn user and item latent factors from user-item historic behaviors. To apply it into real big data scenarios, efficiency becomes the first concern, including offline model training efficiency and online recommendation efficiency. In this paper, we propose a Distributed Collaborative Hashing (DCH) model which can significantly improve both efficiencies. Specifically, we first propose a distributed learning framework, following the state-of-the-art parameter server paradigm, to learn the offline collaborative model. Our model can be learnt efficiently by distributedly computing subgradients in minibatches on workers and updating model parameters on servers asynchronously. We then adopt hashing technique to speedup the online recommendation procedure. Recommendation can be quickly made through exploiting lookup hash tables. We conduct thorough experiments on two real large-scale datasets. The experimental results demonstrate that, comparing with the classic and state-of-the-art (distributed) latent factor models, DCH has comparable performance in terms of recommendation accuracy but has both fast convergence speed in offline model training procedure and realtime efficiency in online recommendation procedure. Furthermore, the encouraging performance of DCH is also shown for several real-world applications in Ant Financial.

cs.LG↗

arXiv2018

A Boosting Framework of Factorization Machine

Recently, Factorization Machines (FM) has become more and more popular for recommendation systems, due to its effectiveness in finding informative interactions between features. Usually, the weights for the interactions is learnt as a low rank weight matrix, which is formulated as an inner product of two low rank matrices. This low rank can help improve the generalization ability of Factorization Machines. However, to choose the rank properly, it usually needs to run the algorithm for many times using different ranks, which clearly is inefficient for some large-scale datasets. To alleviate this issue, we propose an Adaptive Boosting framework of Factorization Machines (AdaFM), which can adaptively search for proper ranks for different datasets without re-training. Instead of using a fixed rank for FM, the proposed algorithm will adaptively gradually increases its rank according to its performance until the performance does not grow, using boosting strategy. To verify the performance of our proposed framework, we conduct an extensive set of experiments on many real-world datasets. Encouraging empirical results shows that the proposed algorithms are generally more effective than state-of-the-art other Factorization Machines.

cs.LG↗

arXiv2018

Machine Learning-Assisted Least Loaded Routing to Improve Performance of Circuit-Switched Networks

The Least Loaded (LL) routing algorithm has been in recent decades the routing method of choice in circuit switched networks and therefore it provides a benchmark against which new methods can be compared. This paper improves the performance of the LL algorithm by additionally incorporating a machine learning approach, using a conceptually simple supervised naïve Bayes (NB) classifier. Based on a sequence of historical network snapshots, this predicts the potential future circuit blocking probability between each node pair. These snapshots are taken for each service request arriving to the network and record the number of busy capacity units on each link at that instant. The candidate route for serving a current service request is based on both the link loads and the potential future blocking probability of the entire network in case this route is indeed used. The performance of this proposed approach is studied via simulations and compared with both the conventional LL algorithm and the Shortest Path (SP) based approach. Results indicate that the proposed supervised naïve Bayes classifier-assisted LL routing algorithm significantly reduces blocking probability of service connection requests and outperforms both the conventional LL and SP routing algorithms. To enable the learning process based on a large number of network snapshots, we also develop a parallel computing framework to implement parallel learning and performance evaluation. Also, a network control system supporting naïve Bayes classifier-assisted LL routing algorithm is addressed.

cs.NI↗

arXiv2017

Numerical methods for thermally stressed shallow shell equations

We develop efficient and accurate numerical methods to solve a class of shallow shell problems of the von Karman type. The governing equations form a fourth-order coupled system of nonlinear biharnomic equations for the transverse deflection and Airy's stress function. A second-order finite difference discretization with three iterative methods (Picard, Newton and Trust-Region Dogleg) are proposed for the numerical solution of the nonlinear PDE system. Three simple boundary conditions and two application-motivated mixed boundary conditions are considered. Along with the nonlinearity of the system, boundary singularities that appear when mixed boundary conditions are specified are the main numerical challenges. Two approaches that use either a transition function or local corrections are developed to deal with these boundary singularities. All the proposed numerical methods are validated using carefully designed numerical tests, where expected orders of accuracy and rates of convergence are observed. A rough run-time performance comparison is also conducted to illustrate the efficiency of our methods. As an application of the methods, a snap-through thermal buckling problem is considered. The critical thermal loads of shell buckling with various boundary conditions are numerically calculated, and snap-through bifurcation curves are also obtained using our numerical methods together with a pseudo-arclength continuation method. Our results are consistent with previous studies.

math.NA↗

arXiv2016

Computed Flow and Fluorescence Over the Ocular Surface

Fluorescein is perhaps the most commonly used substance to visualize tear film thickness and dynamics; better understanding of this process aids understanding of dry eye syndrome which afflicts millions of people. We study a mathematical model for tear film flow, evaporation, solutal transport and fluorescence over the exposed ocular surface during the interblink. Transport of the fluorescein ion by fluid flow in the tear film affects the intensity of fluorescence via changes in concentration and tear film thickness. Evaporation causes increased osmolarity and potential irritation over the ocular surface; it also alters fluorescein concentration and thus fluorescence. Using thinning rates from in vivo measurements together with thin film equations for flow and transport of multiple solutes, we compute dynamic results for tear film quantities of interest. We compare our computed intensity distributions with in vivo observations. A number of experimental features are recovered by the model.

q-bio.QM↗