arXiv subjects

Qizhang Li

Publications and source records attributed to Qizhang Li.

25 records · Page 2Linked to original sources

arXiv2022

On Steering Multi-Annotations per Sample for Multi-Task Learning

The study of multi-task learning has drawn great attention from the community. Despite the remarkable progress, the challenge of optimally learning different tasks simultaneously remains to be explored. Previous works attempt to modify the gradients from different tasks. Yet these methods give a subjective assumption of the relationship between tasks, and the modified gradient may be less accurate. In this paper, we introduce Stochastic Task Allocation~(STA), a mechanism that addresses this issue by a task allocation approach, in which each sample is randomly allocated a subset of tasks. For further progress, we propose Interleaved Stochastic Task Allocation~(ISTA) to iteratively allocate all tasks to each example during several consecutive iterations. We evaluate STA and ISTA on various datasets and applications: NYUv2, Cityscapes, and COCO for scene understanding and instance segmentation. Our experiments show both STA and ISTA outperform current state-of-the-art methods. The code will be available.

cs.CV↗

arXiv2021

Giant radiative thermal rectification using an intrinsic semiconductor film

Rectification of heat flow via a thermal diode is not only of fundamental interest, but can also enable a range of novel applications in thermal management and energy conversion. However, despite decades of extensive research, large rectification ratios of practical importance have yet to be demonstrated. Here, we theoretically achieve giant rectification ratios (3 to almost 5 orders of magnitude) by leveraging near-field radiative thermal transport between two parallel planes. Guided by a rational design approach centering on the electromagnetic local density of states (LDOS), we employ a thin film of an intrinsic semiconductor-such as silicon-as one terminal of our radiative thermal diodes, which provides the necessary nonlinearity and a substantial LDOS contrast as the temperature bias is flipped. For the other terminal, we explore two kinds of materials which either serve as a narrowband or a broadband filter, both capable of converting the large LDOS contrast into giant thermal rectification. We further consider representative multilayer configurations in order to block backside radiation and improve mechanical stability. All the diodes perform well over a wide range of film thicknesses, gap sizes, and temperatures. Our work offers an opportunity for realizing thermal diodes with unprecedented rectification ratios.

cond-mat.mes-hall↗

arXiv2021

Radiative thermal diode via hyperbolic metamaterials

Hyperbolic metamaterials (HMMs) support propagating waves with arbitrarily large wavevectors over broad spectral ranges, and are uniquely valuable for engineering radiative thermal transport in the near field. Here, by employing a rational design approach based on the electromagnetic local density of states, we demonstrate the ability of HMMs to substantially rectify radiative heat flow. Our idea is to establish a forward-biased scenario where the two HMM-based terminals of a thermal diode feature overlapped hyperbolic bands which result in a large heat current, and suppress the reverse heat flow by creating spectrally mismatched density of states as the temperature bias is flipped. As an example, we present a few high-performance thermal diodes by pairing HMMs made of polar dielectrics and metal-to-insulator transition (MIT) materials in the form of periodic nanowire arrays, and considering three representative kinds of substrates. Upon optimization, we theoretically achieve a rectification ratio of 324 at a 100 nm gap, which remains greater than 148 for larger gap sizes up to 1 um over a wide temperature range. The maximum rectification represents an almost 1000-fold increase compared to a bulk diode using the same materials, and is twice that of state-of-the-art designs. Our work highlights the potential of HMMs for rectifying radiative heat flow, and may find applications in advanced thermal management and energy conversion systems.

physics.optics↗

arXiv2021

Thin-film radiative thermal diode with large rectification

We propose a mechanism to substantially rectify radiative heat flow by matching thin films of metal-to-insulator transition materials and polar dielectrics in the electromagnetic near field. By leveraging the distinct scaling behaviors of the local density of states with film thickness for metals and insulators, we theoretically achieve rectification ratios over 140-a 10-fold improvement over the state of the art-with nanofilms of vanadium dioxide and cubic boron nitride in the parallel-plane geometry at experimentally feasible gap sizes (~100 nm). Our rational design offers relative ease of fabrication, flexible choice of materials, and robustness against deviations from optimal film thicknesses. We expect this work to facilitate the application of thermal diodes in solid-state thermal circuits and energy conversion devices.

physics.app-ph↗

arXiv2020

Backpropagating Linearly Improves Transferability of Adversarial Examples

The vulnerability of deep neural networks (DNNs) to adversarial examples has drawn great attention from the community. In this paper, we study the transferability of such examples, which lays the foundation of many black-box attacks on DNNs. We revisit a not so new but definitely noteworthy hypothesis of Goodfellow et al.'s and disclose that the transferability can be enhanced by improving the linearity of DNNs in an appropriate manner. We introduce linear backpropagation (LinBP), a method that performs backpropagation in a more linear fashion using off-the-shelf attacks that exploit gradients. More specifically, it calculates forward as normal but backpropagates loss as if some nonlinear activations are not encountered in the forward pass. Experimental results demonstrate that this simple yet effective method obviously outperforms current state-of-the-arts in crafting transferable adversarial examples on CIFAR-10 and ImageNet, leading to more effective attacks on a variety of DNNs.

cs.LG↗

arXiv2020

Practical No-box Adversarial Attacks against DNNs

The study of adversarial vulnerabilities of deep neural networks (DNNs) has progressed rapidly. Existing attacks require either internal access (to the architecture, parameters, or training set of the victim model) or external access (to query the model). However, both the access may be infeasible or expensive in many scenarios. We investigate no-box adversarial examples, where the attacker can neither access the model information or the training set nor query the model. Instead, the attacker can only gather a small number of examples from the same problem domain as that of the victim model. Such a stronger threat model greatly expands the applicability of adversarial attacks. We propose three mechanisms for training with a very small dataset (on the order of tens of examples) and find that prototypical reconstruction is the most effective. Our experiments show that adversarial examples crafted on prototypical auto-encoding models transfer well to a variety of image classification and face verification models. On a commercial celebrity recognition system held by clarifai.com, our approach significantly diminishes the average prediction accuracy of the system to only 15.40%, which is on par with the attack that transfers adversarial examples from a pre-trained Arcface model.

cs.CV↗

arXiv2020

Yet Another Intermediate-Level Attack

The transferability of adversarial examples across deep neural network (DNN) models is the crux of a spectrum of black-box attacks. In this paper, we propose a novel method to enhance the black-box transferability of baseline adversarial examples. By establishing a linear mapping of the intermediate-level discrepancies (between a set of adversarial inputs and their benign counterparts) for predicting the evoked adversarial loss, we aim to take full advantage of the optimization procedure of multi-step baseline attacks. We conducted extensive experiments to verify the effectiveness of our method on CIFAR-100 and ImageNet. Experimental results demonstrate that it outperforms previous state-of-the-arts considerably. Our code is at https://github.com/qizhangli/ila-plus-plus.

cs.CV↗