arXiv subjects

Minh Pham

Publications and source records attributed to Minh Pham.

34 records · Page 2Linked to original sources

arXiv2022

Revisiting Self-Distillation

Knowledge distillation is the procedure of transferring "knowledge" from a large model (the teacher) to a more compact one (the student), often being used in the context of model compression. When both models have the same architecture, this procedure is called self-distillation. Several works have anecdotally shown that a self-distilled student can outperform the teacher on held-out data. In this work, we systematically study self-distillation in a number of settings. We first show that even with a highly accurate teacher, self-distillation allows a student to surpass the teacher in all cases. Secondly, we revisit existing theoretical explanations of (self) distillation and identify contradicting examples, revealing possible drawbacks of these explanations. Finally, we provide an alternative explanation for the dynamics of self-distillation through the lens of loss landscape geometry. We conduct extensive experiments to show that self-distillation leads to flatter minima, thereby resulting in better generalization.

cs.LG↗

arXiv2022

Transformer with Fourier Integral Attentions

Multi-head attention empowers the recent success of transformers, the state-of-the-art models that have achieved remarkable success in sequence modeling and beyond. These attention mechanisms compute the pairwise dot products between the queries and keys, which results from the use of unnormalized Gaussian kernels with the assumption that the queries follow a mixture of Gaussian distribution. There is no guarantee that this assumption is valid in practice. In response, we first interpret attention in transformers as a nonparametric kernel regression. We then propose the FourierFormer, a new class of transformers in which the dot-product kernels are replaced by the novel generalized Fourier integral kernels. Different from the dot-product kernels, where we need to choose a good covariance matrix to capture the dependency of the features of data, the generalized Fourier integral kernels can automatically capture such dependency and remove the need to tune the covariance matrix. We theoretically prove that our proposed Fourier integral kernels can efficiently approximate any key and query distributions. Compared to the conventional transformers with dot-product attention, FourierFormers attain better accuracy and reduce the redundancy between attention heads. We empirically corroborate the advantages of FourierFormers over the baseline transformers in a variety of practical applications including language modeling and image classification.

cs.LG↗

arXiv2022

Smooth-Reduce: Leveraging Patches for Improved Certified Robustness

Randomized smoothing (RS) has been shown to be a fast, scalable technique for certifying the robustness of deep neural network classifiers. However, methods based on RS require augmenting data with large amounts of noise, which leads to significant drops in accuracy. We propose a training-free, modified smoothing approach, Smooth-Reduce, that leverages patching and aggregation to provide improved classifier certificates. Our algorithm classifies overlapping patches extracted from an input image, and aggregates the predicted logits to certify a larger radius around the input. We study two aggregation schemes -- max and mean -- and show that both approaches provide better certificates in terms of certified accuracy, average certified radii and abstention rates as compared to concurrent approaches. We also provide theoretical guarantees for such certificates, and empirically show significant improvements over other randomized smoothing methods that require expensive retraining. Further, we extend our approach to videos and provide meaningful certificates for video classifiers. A project page can be found at https://nyu-dice-lab.github.io/SmoothReduce/

cs.LG↗

arXiv2021

Harnessing Geometric Constraints from Emotion Labels to improve Face Verification

For the task of face verification, we explore the utility of harnessing auxiliary facial emotion labels to impose explicit geometric constraints on the embedding space when training deep embedding models. We introduce several novel loss functions that, in conjunction with a standard Triplet Loss [43], or ArcFace loss [10], provide geometric constraints on the embedding space; the labels for our loss functions can be provided using either manually annotated or automatically detected auxiliary emotion labels. Our method is implemented purely in terms of the loss function and does not require any changes to the neural network backbone of the embedding function.

cs.CV↗

arXiv2021

Enhanced lasing action and spontaneous emission in periodic plasmonic nanohole arrays

Periodic arrays of air nanoholes in thin metal films that support surface plasmon resonances can provide an alternative approach for boosting the light-matter interactions at the nanoscale. Indeed, nanohole arrays have garnered great interest in recent years for their use in biosensing, light emission enhancement and spectroscopy. However, the large-scale use of nanohole arrays in emerging technology requires new low-cost fabrication techniques. Here, we demonstrate a simple technique to fabricate nanohole arrays and examine their photonic applications. In contrast to the complicated and most commonly used electron beam lithography technique, hexagonal arrays of nanoholes are fabricated by using a simple combination of shadowing nanosphere lithography technique and electron beam deposition. These arrays are shown to offer enhancements in the lasing emission of an organic dye liquid gain medium with a quality factor above 150. Additionally, a 7-fold increase in Purcell factor is observed for CdSe quantum dot-integrated nanohole arrays.

physics.optics↗

arXiv2021

Direct observation of 3D topological spin textures and their interactions using soft x-ray vector ptychography

Magnetic topological defects are energetically stable spin configurations characterized by symmetry breaking. Vortices and skyrmions are two well-known examples of 2D spin textures that have been actively studied for both fundamental interest and practical applications. However, experimental evidence of the 3D spin textures has been largely indirect or qualitative to date, due to the difficulty of quantitively characterizing them within nanoscale volumes. Here, we develop soft x-ray vector ptychography to quantitatively image the 3D magnetization vector field in a frustrated superlattice with 10 nm spatial resolution. By applying homotopy theory to the experimental data, we quantify the topological charge of hedgehogs and anti-hedgehogs as emergent magnetic monopoles and probe their interactions inside the frustrated superlattice. We also directly observe virtual hedgehogs and anti-hedgehogs created by magnetically inert voids. We expect that this new quantitative imaging method will open the door to study 3D topological spin textures in a broad class of magnetic materials. Our work also demonstrates that magnetically frustrated superlattices could be used as a new platform to investigate hedgehog interactions and dynamics and to exploit optimized geometries for information storage and transport applications.

cond-mat.str-el↗

arXiv2020

Direct observation of 3D atomic packing in monatomic amorphous materials

Liquids and solids are two fundamental states of matter. However, due to the lack of direct experimental determination, our understanding of the 3D atomic structure of liquids and amorphous solids remained speculative. Here we advance atomic electron tomography to determine for the first time the 3D atomic positions in monatomic amorphous materials, including a Ta thin film and two Pd nanoparticles. We observe that pentagonal bipyramids are the most abundant atomic motifs in these amorphous materials. Instead of forming icosahedra, the majority of pentagonal bipyramids arrange into networks that extend to medium-range scale. Molecular dynamic simulations further reveal that pentagonal bipyramid networks are prevalent in monatomic amorphous liquids, which rapidly grow in size and form icosahedra during the quench from the liquid state to glass state. The experimental method and results are expected to advance the study of the amorphous-crystalline phase transition and glass transition at the single-atom level.

cond-mat.mtrl-sci↗

arXiv2020

Determining the three-dimensional atomic structure of a metallic glass

Amorphous solids such as glass are ubiquitous in our daily life and have found broad applications ranging from window glass and solar cells to telecommunications and transformer cores. However, due to the lack of long-range order, the three-dimensional (3D) atomic structure of amorphous solids have thus far defied any direct experimental determination without model fitting. Here, using a multi-component metallic glass as a proof-of-principle, we advance atomic electron tomography to determine the 3D atomic positions in an amorphous solid for the first time. We quantitatively characterize the short-range order (SRO) and medium-range order (MRO) of the 3D atomic arrangement. We find that although the 3D atomic packing of the SRO is geometrically disordered, some SRO connect with each other to form crystal-like networks and give rise to MRO. We identify four crystal-like MRO networks - face-centred cubic, hexagonal close-packed, body-centered cubic and simple cubic - coexisting in the sample, which show translational but no orientational order. These observations confirm that the 3D atomic structure in some parts of the sample is consistent with the efficient cluster packing model. Looking forward, we anticipate this experiment will open the door to determining the 3D atomic coordinates of various amorphous solids, whose impact on non-crystalline solids may be comparable to the first 3D crystal structure solved by x-ray crystallography over a century ago.

cond-mat.dis-nn↗

arXiv2020

A Survey on Security Attacks and Defense Techniques for Connected and Autonomous Vehicles

Autonomous Vehicle has been transforming intelligent transportation systems. As telecommunication technology improves, autonomous vehicles are getting connected to each other and to infrastructures, forming Connected and Autonomous Vehicles (CAVs). CAVs will help humans achieve safe, efficient, and autonomous transportation systems. However, CAVs will face significant security challenges because many of their components are vulnerable to attacks, and a successful attack on a CAV may have significant impacts on other CAVs and infrastructures due to their communications. In this paper, we conduct a survey on 184 papers from 2000 to 2020 to understand state-of-the-art CAV attacks and defense techniques. This survey first presents a comprehensive overview of security attacks and their corresponding countermeasures on CAVs. We then discuss the details of attack models based on the targeted CAV components of attacks, access requirements, and attack motives. Finally, we identify some current research challenges and trends from the perspectives of both academic research and industrial development. Based on our studies of academic literature and industrial publications, we have not found any strong connection between academic research and industry's implementation on CAV-related security issues. While efforts from CAV manufacturers to secure CAVs have been reported, there is no evidence to show that CAVs on the market have the ability to defend against some novel attack models that the research community has recently found. This survey may give researchers and engineers a better understanding of the current status and trend of CAV security for CAV future improvement.

cs.CR↗

arXiv2020

RESIRE: real space iterative reconstruction engine for Tomography

Tomography has made a revolutionary impact on diverse fields, ranging from macro-/mesoscopic scale studies in biology, radiology, plasma physics to the characterization of 3D atomic structure in material science. The fundamental of tomography is to reconstruct a 3D object from a set of 2D projections. To solve the tomography problem, many algorithms have been developed. Among them are methods using transformation technique such as computed tomography (CT) based on Radon transform and Generalized Fourier iterative reconstruction (GENFIRE) based on Fourier slice theorem (FST), and direct methods such as Simultaneous Iterative Reconstruction Technique (SIRT) and Simultaneous Algebraic Reconstruction Technique (SART) using gradient descent and algebra technique. In this paper, we propose a hybrid gradient descent to solve the tomography problem by combining Fourier slice theorem and calculus of variations. By using simulated and experimental data, we show that the state-of-art RESIRE can produce more superior results than previous methods; the reconstructed objects have higher quality and smaller relative errors. More importantly, RESIRE can deal with partially blocked projections rigorously where only part of projection information are provided while other methods fail. We anticipate RESIRE will not only improve the reconstruction quality in all existing tomographic applications, but also expand tomography method to a broad class of functional thin films. We expect RESIRE to find a broad applications across diverse disciplines.

math.OC↗

arXiv2020

Potential of attosecond coherent diffractive imaging

Attosecond science has been transforming our understanding of electron dynamics in atoms, molecules and solids. However, to date almost all of the attoscience experiments have been based on spectroscopic measurements because attosecond pulses have intrinsically very broad spectra due to the uncertainty principle and are incompatible with conventional imaging systems. Here we report an important advance towards achieving attosecond coherent diffractive imaging. Using simulated attosecond pulses, we simultaneously reconstruct the spectrum, 17 probes and 17 spectral images of extended objects from a set of ptychographic diffraction patterns. We further confirm the principle and feasibility of this method by successfully performing a ptychographic coherent diffractive imaging experiment using a light-emitting diode with a broad spectrum. We believe this work clear the way to an unexplored domain of attosecond imaging science, which could have a far-reaching impact across different disciplines.

physics.optics↗

arXiv2019

A semi-implicit relaxed Douglas-Rachford algorithm (sir-DR) for Ptychograhpy

Alternating projection based methods, such as ePIE and rPIE, have been used widely in ptychography. However, they only work well if there are adequate measurements (diffraction patterns); in the case of sparse data (i.e. fewer measurements) alternating projection underperforms and might not even converge. In this paper, we propose semi-implicit relaxed Douglas Rachford (sir-DR), an accelerated iterative method, to solve the classical ptychography problem. Using both simulated and experimental data, we show that sir-DR improves the convergence speed and the reconstruction quality relative to ePIE and rPIE. Furthermore, in certain cases when sparsity is high, sir-DR converges while ePIE and rPIE fail. To facilitate others to use the algorithm, we post the Matlab source code of sir-DR on a public website (www.physics.ucla.edu/research/imaging/sir-DR). We anticipate that this algorithm can be generally applied to the ptychographic reconstruction of a wide range of samples in the physical and biological sciences.

eess.IV↗

arXiv2019

Laplacian Smoothing Gradient Descent

We propose a class of very simple modifications of gradient descent and stochastic gradient descent. We show that when applied to a large variety of machine learning problems, ranging from logistic regression to deep neural nets, the proposed surrogates can dramatically reduce the variance, allow to take a larger step size, and improve the generalization accuracy. The methods only involve multiplying the usual (stochastic) gradient by the inverse of a positive definitive matrix (which can be computed efficiently by FFT) with a low condition number coming from a one-dimensional discrete Laplacian or its high order generalizations. It also preserves the mean and increases the smallest component and decreases the largest component. The theory of Hamilton-Jacobi partial differential equations demonstrates that the implicit version of the new algorithm is almost the same as doing gradient descent on a new function which (i) has the same global minima as the original function and (ii) is ``more convex". Moreover, we show that optimization algorithms with these surrogates converge uniformly in the discrete Sobolev $H_σ^p$ sense and reduce the optimality gap for convex optimization problems. The code is available at: \url{https://github.com/BaoWangMath/LaplacianSmoothing-GradientDescent}

cs.LG↗

arXiv2018

Stochastic Backward Euler: An Implicit Gradient Descent Algorithm for $k$-means Clustering

In this paper, we propose an implicit gradient descent algorithm for the classic $k$-means problem. The implicit gradient step or backward Euler is solved via stochastic fixed-point iteration, in which we randomly sample a mini-batch gradient in every iteration. It is the average of the fixed-point trajectory that is carried over to the next gradient step. We draw connections between the proposed stochastic backward Euler and the recent entropy stochastic gradient descent (Entropy-SGD) for improving the training of deep neural networks. Numerical experiments on various synthetic and real datasets show that the proposed algorithm provides better clustering results compared to $k$-means algorithms in the sense that it decreased the objective function (the cluster) and is much more robust to initialization.

math.OC↗

arXiv2018

Generalized Proximal Smoothing for Phase Retrieval

In this paper, we report the development of the generalized proximal smoothing (GPS) algorithm for phase retrieval of noisy data. GPS is a optimization-based algorithm, in which we relax both the Fourier magnitudes and object constraints. We relax the object constraint by introducing the generalized Moreau-Yosida regularization and heat kernel smoothing. We are able to readily handle the associated proximal mapping in the dual variable by using an infimal convolution. We also relax the magnitude constraint into a least squares fidelity term, whose proximal mapping is available. GPS alternatively iterates between the two proximal mappings in primal and dual spaces, respectively. Using both numerical simulation and experimental data, we show that GPS algorithm consistently outperforms the classical phase retrieval algorithms such as hybrid input-output (HIO) and oversampling smoothness (OSS), in terms of the convergence speed, consistency of the phase retrieval, and robustness to noise.

math.OC↗

arXiv2014

Alternating Linearization for Structured Regularization Problems

We adapt the alternating linearization method for proximal decomposition to structured regularization problems, in particular, to the generalized lasso problems. The method is related to two well-known operator splitting methods, the Douglas--Rachford and the Peaceman--Rachford method, but it has descent properties with respect to the objective function. This is achieved by employing a special update test, which decides whether it is beneficial to make a Peaceman--Rachford step, any of the two possible Douglas--Rachford steps, or none. The convergence mechanism of the method is related to that of bundle methods of nonsmooth optimization. We also discuss implementation for very large problems, with the use of specialized algorithms and sparse data structures. Finally, we present numerical results for several synthetic and real-world examples, including a three-dimensional fused lasso problem, which illustrate the scalability, efficacy, and accuracy of the method.

stat.CO↗