arXiv subjects

Tianyuan Zhang

Publications and source records attributed to Tianyuan Zhang.

61 records · Page 4Linked to original sources

arXiv2021

DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries

We introduce a framework for multi-camera 3D object detection. In contrast to existing works, which estimate 3D bounding boxes directly from monocular images or use depth prediction networks to generate input for 3D object detection from 2D information, our method manipulates predictions directly in 3D space. Our architecture extracts 2D features from multiple camera images and then uses a sparse set of 3D object queries to index into these 2D features, linking 3D positions to multi-view images using camera transformation matrices. Finally, our model makes a bounding box prediction per object query, using a set-to-set loss to measure the discrepancy between the ground-truth and the prediction. This top-down approach outperforms its bottom-up counterpart in which object bounding box prediction follows per-pixel depth estimation, since it does not suffer from the compounding error introduced by a depth prediction model. Moreover, our method does not require post-processing such as non-maximum suppression, dramatically improving inference speed. We achieve state-of-the-art performance on the nuScenes autonomous driving benchmark.

cs.CV↗

arXiv2021

Efficient Four-Component Dirac-Coulomb-Gaunt Hartree--Fock in Pauli Spinor Representation

Four-component Dirac Hartree--Fock is an accurate mean-field method for treating molecular systems where relativistic effects are important. However, the computational cost and complexity of the two-electron interaction makes this method less common, even though we can consider the Dirac Hartree--Fock Hamiltonian the "ground truth" of electronic structure, barring explicit quantum-electrodynamical effects. Being able to calculate these effects is then vital to the design of lower scaling methods for accurate predictions in computational spectroscopy and properties of heavy element complexes that must include relativistic effects for even qualitative accuracy. In this work, we present a Pauli quaternion formalism of maximal component- and spin-separation for computing the Dirac-Coulomb-Gaunt Hartree--Fock ground state, with a minimal floating-point-operation count algorithm. This approach also allows one to explicitly separate different spin physics from the two-body interactions, such as spin-free, spin-orbit, and the spin-spin contributions. Additionally, we use this formalism to examine relativistic trends in the periodic table, and analyze the basis set dependence of atomic gold and gold dimer systems.

physics.chem-ph↗

arXiv2019

Domain-Aware Dynamic Networks

Deep neural networks with more parameters and FLOPs have higher capacity and generalize better to diverse domains. But to be deployed on edge devices, the model's complexity has to be constrained due to limited compute resource. In this work, we propose a method to improve the model capacity without increasing inference-time complexity. Our method is based on an assumption of data locality: for an edge device, within a short period of time, the input data to the device are sampled from a single domain with relatively low diversity. Therefore, it is possible to utilize a specialized, low-complexity model to achieve good performance in that input domain. To leverage this, we propose Domain-aware Dynamic Network (DDN), which is a high-capacity dynamic network in which each layer contains multiple weights. During inference, based on the input domain, DDN dynamically combines those weights into one single weight that specializes in the given domain. This way, DDN can keep the inference-time complexity low but still maintain a high capacity. Experiments show that without increasing the parameters, FLOPs, and actual latency, DDN achieves up to 2.6\% higher AP50 than a static network on the BDD100K object-detection benchmark.

cs.LG↗

arXiv2019

You Only Propagate Once: Accelerating Adversarial Training via Maximal Principle

Deep learning achieves state-of-the-art results in many tasks in computer vision and natural language processing. However, recent works have shown that deep networks can be vulnerable to adversarial perturbations, which raised a serious robustness issue of deep networks. Adversarial training, typically formulated as a robust optimization problem, is an effective way of improving the robustness of deep networks. A major drawback of existing adversarial training algorithms is the computational overhead of the generation of adversarial examples, typically far greater than that of the network training. This leads to the unbearable overall computational cost of adversarial training. In this paper, we show that adversarial training can be cast as a discrete time differential game. Through analyzing the Pontryagin's Maximal Principle (PMP) of the problem, we observe that the adversary update is only coupled with the parameters of the first layer of the network. This inspires us to restrict most of the forward and back propagation within the first layer of the network during adversary updates. This effectively reduces the total number of full forward and backward propagation to only one for each group of adversary updates. Therefore, we refer to this algorithm YOPO (You Only Propagate Once). Numerical experiments demonstrate that YOPO can achieve comparable defense accuracy with approximately 1/5 ~ 1/4 GPU time of the projected gradient descent (PGD) algorithm. Our codes are available at https://https://github.com/a1600012888/YOPO-You-Only-Propagate-Once.

stat.ML↗

arXiv2019

Interpreting Adversarially Trained Convolutional Neural Networks

We attempt to interpret how adversarially trained convolutional neural networks (AT-CNNs) recognize objects. We design systematic approaches to interpret AT-CNNs in both qualitative and quantitative ways and compare them with normally trained models. Surprisingly, we find that adversarial training alleviates the texture bias of standard CNNs when trained on object recognition tasks, and helps CNNs learn a more shape-biased representation. We validate our hypothesis from two aspects. First, we compare the salience maps of AT-CNNs and standard CNNs on clean images and images under different transformations. The comparison could visually show that the prediction of the two types of CNNs is sensitive to dramatically different types of features. Second, to achieve quantitative verification, we construct additional test datasets that destroy either textures or shapes, such as style-transferred version of clean data, saturated images and patch-shuffled ones, and then evaluate the classification accuracy of AT-CNNs and normal CNNs on these datasets. Our findings shed some light on why AT-CNNs are more robust than those normally trained ones and contribute to a better understanding of adversarial training over CNNs from an interpretation perspective.

cs.LG↗

arXiv2019

Improving the efficiency of the multireference driven similarity renormalization group via sequential transformation, density fitting, and the non-interacting virtual orbital approximation

This study examines several techniques to improve the efficiency of the linearized multireference driven similarity renormalization group truncated to one- and two-body operators [MR-LDSRG(2)]. We propose a sequential MR-LDSRG(2) [sq-MR-LDSRG(2)] approach, in which one-body rotations are folded exactly into the Hamiltonian. This new approach is combined with density fitting (DF) to reduce the storage cost of two-electron integrals. To further avoid the storage of large four-index intermediates, we propose a non-interacting virtual orbit (NIVO) approximation in which tensor blocks labeled by three and four virtual indices are neglected. The NIVO approximation reduces the computational cost prefactor of the MR-LDSRG(2) bringing it closer to that of coupled cluster with singles and doubles (CCSD). We test the effect of the DF and NIVO approximations on the MR-LDSRG(2) and sq-MR-LDSRG(2) methods by computing properties of eight diatomic molecules. The diatomic constants obtained by DF-sq-MR-LDSRG(2)+NIVO are found to be as accurate as those from the original MR-LDSRG(2) and coupled cluster theory with singles, doubles, and perturbative triples. Finally, we demonstrate that the DF-sq-MR-LDSRG(2)+NIVO scheme can be applied to chemical systems with more than 550 basis functions by computing the automerization energy of cyclobutadiene with a quintuple-zeta basis set. The predicted automerization energy is found similar to the values computed with Mukherjee's state-specific multireference coupled cluster theory with singles and doubles.

physics.chem-ph↗

arXiv2016

A deterministic projector configuration interaction approach for the ground state of quantum many-body systems

In this work we propose a novel approach to solve the Schrödinger equation which combines projection onto the ground state with a path-filtering truncation scheme. The resulting projector configuration interaction (PCI) approach realizes a deterministic version of the full configuration interaction quantum Monte Carlo (FCIQMC) method [Booth, G. H.; Thom, A. J. W.; Alavi, A. J. Chem. Phys. 2009, 131, 054106]. To improve upon the linearized imaginary-time propagator, we develop an optimal projector scheme based on an exponential Chebyshev expansion in the limit of an infinite imaginary time step. After writing the exact projector as a path integral in determinant space, we introduce a path filtering procedure that truncates the size of the determinantal basis and approximates the Hamiltonian. The path filtering procedure is controlled by one real threshold that determines the accuracy of the PCI energy and is not biased towards any determinant. Therefore, the PCI approach can equally well describe static and dynamic electron correlation. This point is illustrated in benchmark computation on N$_2$ at both equilibrium and stretched geometries. In both cases, the PCI achieves chemical accuracy with wave functions that contain less than 0.5% of the full CI space. We also report computations on the ground state of C$_2$ with up to quaduple-$ζ$ basis sets and wave functions as large as 200 million determinants, which allow a direct comparison of the PCI, FCIQMC, and density matrix renormalization group (DMRG) methods. The size of the PCI wave function grows modestly with the number of unoccupied orbitals and its accuracy may be tuned to match that of FCIQMC and DMRG.

physics.chem-ph↗