Glossary Collection

目录

• 直接修饰用
• 间接强调用 (多为副词)
• 过渡用
• 特别的名词
• 动词
• 词组
• 各种介词
• 句子
  • 摘要
  • 引言
  • 总结
  • 正文
  • 实验

直接修饰用

Word	含义	例句	近义词
nuanced	adj. 微妙的；具有细微差别的; v. 精确细腻地表演；细致入微地描绘	However, a more nuanced understanding of images arguably requires the ability to reason about how the scene depicted in the image would change in response to interventions.	subtle
poorly	adv. 贫穷地;贫乏地;不充分地; adj. 身体不舒服的; 心情恶劣的	The company has been performing poorly over the past year. poorly defined	deficiently
hazardous	adj. 有危险的；冒险的；碰运气的	Without proper safeguards, users of DNN-based applications can be exposed to unforeseen hazardous situations caused by "trivial" noises.	risky
problematic	adj. 问题的；有疑问的；不确定的	This is problematic if a test sequence is longer than any of the training sequences.	uncertain
problem-oriented	adj. 面向问题的	because it established at the outset that enhancement techniques are problem-oriented ...	-
conceptually	adv. 概念上	We propose the MLP-Mixer architecture (or “Mixer” for short), a competitive but conceptually and technically simple alternative.	-
technically	adv. 技术上；专门地；学术上；工艺上	We propose the MLP-Mixer architecture (or “Mixer” for short), a competitive but conceptually and technically simple alternative.	-
plausible	adj. 看似可信的	It may still be possible to determine a plausible causal structure in practice. Our approach seeks sets of unobserved and plausible hard-examples based on existing training samples.	-
superior	adj. 上级的；优秀的，出众的；高傲的;	In this aspect, content filtering is superior.	surpass; exceed
adequate	adj. 充足的；适当的；胜任的	The structure would not be adequate ... She didn't give an adequate explanation for being late.	enough sufficient; substantial
substantial	adj. 大量的；实质的；内容充实的	Most deep learning methods require substantial amounts of manually labeled data, which restricts their applicability in may domains that suffer from a dearth of annotated resources.	enough; sufficient; adequate
ubiquitous	adj. 普遍存在的；无所不在的	Open-domain dialogue generation, due to tis potential applications, is becoming ubiquitous in the community of natural language processing.	universal; widespread
scarce	adj. 缺乏的，不足的；稀有的 adv. 仅仅；几乎不；几乎没有	Although large unlabeled text corpora are abundant, labeled data for learning theses specific tasks is scarce, making it challenging for discriminatively trained models to perform adequately. However, the observational data of a population may be scarce in practice.	rare; dearth
dearth	n. 缺乏；饥馑；粮食不足	Most deep learning methods require substantial amounts of manually labeled data, which restricts their applicability in may domains that suffer from a dearth of annotated resources.	scarce
concurrent	adj. 并发的;一致的;同时发生的;并存的; n. [数]共点;同时发生的事件	a concurrent work by ...	alongside
subsequent	adj. 随后的	Subsequent experiments failed to replicate these findings.	-
timely	adj. 及时的；适时的 adv. 及时地；早	Therefore, timely diagnosis of the failures is very important to prevent extreme costs for operations and maintenance.	-
resultant	adj. 由此导致的，因而发生的	The unmatched treated subject would then be excluded from the resultant matched sample.	resulting; consequent
intrinsic	adj. 本质的，固有的	This formula shows the intrinsic beauty of mathematics.	inherent
inherent	adj. 固有的；内在的；与生俱来的，遗传的	The inherent weakness of the new diagnostic test is that it cannot detect cancers caused by viruses.	intrinsic
harmonized	adj. 调和	These two should be harmonized.	compatible
needless	adj. 不必要的, 不需要的;	unsafe ML systems may result in needless loss of life ...	unnecessary
aptly	adv. 恰当地, 贴切地	these models are unable to to aptly describe structure in rich datasets.	properly

间接强调用 (多为副词)

Word	含义	例句	近义词
arguably	adv. 可论证地; 可争辩地;可以说是	Venus was arguably a cozier habitat for life than earth.	-
fundamentally	adv. 根本地; 从根本上; 基础地	They remained fundamentally opposed to the plan.	nearly; almost
primarily	adv. 主要地，首要地，根本地	Data augmentation primarily helps combat texture bias as with ImageNet-R.	mainly, fundamentally
rather	adv. 宁可，宁愿；相当；准确地说；或多或少地	-	pretty (口语) =rather (吃惊)>quite (正常)> fairly (略带失望)
adequately	adv. 充分地；足够地；适当地	I describe balance diagnostics for examining whether the propensity score model has been adequately specified.	enough; fully
extensively	adv. 广阔地；广大地; 普遍地	It has been extensively studied that the convergence to a flat region whose curvature is small benefits the generalization of neural networks.	widely; generally
readily	adv. 容易地；乐意地；无困难地	These methods can be readily adapted to stratification on the propensity score and IPTW using the propensity score.	easily
thorough	adj. 彻底的；十分的；周密的	However, a thorough examination of the comparability ...	thoroughly; drastically
invariably	adv. 总是；不变地；一定地	The propensity score matched sample is almost invariably smaller than the original sample.	constantly; unchangeably
solely	adv. 单独地；仅仅	we show that the optimal rank-one strategy depends solely on then energy budget and the smallest singular value of the data matrix	only
therein	adv. 在其中；在那里	Finally, we refer readers to [JNM20] and the references mentioned therein for an empirical study of ...	here|there| where by|of|in|at|from|on|upon|with
by no means	决不; 并没有; 绝不可能	It is by no means clear that this assumption leads to approximation without the curse of dimensionality ...	-

过渡用

Word	含义	例句	近义词
needless to say	不必说;不用说	The problem, needless to say, is the cost involved.	-
alongside	prep. 与...同时;在...旁边;与...一起;共存;与...相比; adv. 在旁边	we learn a data representation in lower dimensions as compared to the original covariates, and addresses confounding bias alongside.	concurrent
simultaneously	adv. 同时地	while simultaneously handling confounding bias using a decorrelation loss	meanwhile; concurrently
jointly	adv. 共同的;连带地	Subsequently, in the outcome prediction network, we learn an embedding of high-cardinality and continuous treatments, jointly with the data representation.	conjointly
in turn	反过来；转而；轮流，依次	These feelings, in turn, can lead to aggressive action, but not automatically. The heavy reliance on BN’s effectiveness to train models in turn prohibits people from exploring higher-capacity models that would be limited by memory.	-
whereas	conj. 然而；鉴于	Some of the studies show positive results, whereas others do not.	but
albeit	conj. 虽然，尽管	but the classifiers have non-trivial resistance to \(\ell_{\infty}\) attacks (albeit, much lower than AdA trained models) ...	though; although
in contrast	与此相反；比较起来	In contrast, when estimating the effect on smoking cessation of an information brochure given by family physicians to patients who are current smokers, the ATE may be of greater interest than the ATT.	on the contrary
as opposed to	与…截然相反；对照	and to keep them at test time (as opposed to freeze and simplify them out as done for batch normalization). As opposed to the previous case, now the attacker seeks to fool the system leading it to predict that two face images ...	in contrast to; contrary to; unlike
be disjoint from	与...不相交	The augmentations from AugMix are disjoint from ImageNet-C's corruptions.	be distinct from
-		However, the converse is not true in general.

特别的名词

Word	含义	例句	近义词
manifestation	n. 显示; 证明; 临床表现	All causal relationships are manifestations of the powers of objects.	argumentation; demonstration
argumentation	n. 论证; 观点; 争论	However, he did not have to be careful analysis and argumentation.	manifestation; demonstration
demonstration	n. 示范; 证明; 示威游行	We were given a brief demonstration of the computer's functions.	manifestation; argumentation
paradigm	n. 范例；词形变化表	End-to-end neural models are generally trained in a data-driven paradigm.	manner
budget	n. 预算; vt.安排,预定; v.做预算; adj.廉价的	a perturbation budget	-
tolerance	n. 宽容，容忍；耐力；公差	The non-negative L is a constant parameter that describes the tolerance level of varying age relationship.	-
quadruplet	n. 四胎儿；四胞胎之一；成套的四件	Each time, we sample a quadruplet (i, j, k, l) of facial images as the input of the CNN.	triplet; quint
alternative	n. 二中择一；供替代的选择 adj. 供选择的；选择性的；交替的	We present Group Normalization (GN) as a simple alternative to BN.	substitute proxy
substitute	n. 代用品；代替者 v. 替代	... word-level augmentation enriches the training samples by substituting the words in the original sample.	replace
counterpart	n. 副本；配对物；极相似的人或物	In this paper, we ask whether a DNN can distinguish adversarial samples from their normal and noisy counterparts.
abbreviation	n. 缩写；缩写词	If not further specified, we will use the term network, or the abbreviation NN, to refer to FC neural networks.	-
appeal	n. 呼吁，恳求；上诉，申诉；吸引力，感染力，魅力；启发，打动	A major appeal of collaborative filtering is that it is domain free	advantage

动词

phrase	含义	例句	近义词
circumvent	v. 规避;绕行;包围;智取;	We circumvent this challenge using proxy distributions.	avoid
tackle	v. 应付, 应对, 处理（难题或局面)	Our confidence-calibrated adversarial training (CCAT) tackles this problem by biasing the model towards low confidence predictions on adversarial examples.	address; solve; deal with; handle
accomplish	v. 完成，实现	and several approaches have been proposed to accomplish this.	realize; ac
exacerbate	vt. 使加剧; 使恶化; 激怒	The presence of high-dimensional covariates exacerbates the impact of bias as it is harder to isolate and measure the impact of these confounders.	-
degrade	vt. 贬低；使……丢脸；使……降级；使……降解 vi. 降级，降低；退化	Those FCN-driven approaches have weaknesses that might degrade their performance in practice.	deteriorate
incorporate	vt. 包含, 吸收; 体现; 把...合并 vi.合并;混合 adj. 合并的; 一体化的	In our daily lives, humans often justify several actions and events in terms of cause and effect and incorporate the feedback into future actions in order to alter future outcomes. The proposed method can theoretically incorporate arbitrary FCN-based backbone network	combine integrate
integrate	vt. 使…完整；使…成整体；求…的积分；表示…的总和 vi. 求积分；取消隔离；成为一体 adj. 整合的；完全的 n. 一体化；集成体	Hence, there have been several efforts by researchers to integrate causality into machine learning models for obtaining robust and generalizable machine learning models. Dong et al. proposed an iterative algorithm that integrates a momentum term into the iterative process to boost adversarial attacks.	incorporate combine
combine	vt. 使化合；使联合，使结合 vi. 联合，结合；化合 n. 联合收割机；联合企业	The trip will combine business with pleasure. All of this could combine into a perfect storm shattering European unity.	incorporate; integrate
motivate	v. 使有动机，成为……的动机；给出理由；刺激，激发…的积极性；申请	We motivate the need for finding a low-dimensional representation.	-
benefit	vt. 有益于，对…有益 vi. 受益，得益	It has been extensively studied that the convergence to a flat region whose curvature is small benefits the generalization of neural networks. It is thus a natural question to ask whether robust and catastrophic overfitting are related, and whether GradAlign can be beneficial to mitigate robust overfitting. Then we provide evidence against the explanation given by [47] on the benefit of randomness for FGSM AT.	enjoy; encourage
enrich	vt. 使充实；使肥沃；使富足	Optimize \(\theta\) on \(\mathcal{D}\) enriched with data manipulation.	extend
exploit	vt. 开发，利用，开拓；剥削；开采	A context-aware restoration component exploits low-level similarity between each pixel and its context to refine the saliency scores provided by the backbone network. Indeed, since the FR systems exploit a DNN model to generate face descriptors only ...	develop
employ	vt. 使用，采用；雇用；使忙于，使从事于	... many adversarial training methods employ the most adversarial data maximizing the loss for updating the current model such as standard adversarial training (AT)	exploit; adopt; use
recur to	借助于	To prove that, we need to recur to Lemma 1 based on ...	with the help/aid of
explore	v. 探索；探测；探险	we explore a semi-supervised approach for language understanding tasks ...	-
elaborate	vt. 精心制作；详细阐述；vi. 详细描述；变复杂; adj. 精心制作的；详尽的；煞费苦心的	In this section, we elaborate the details of our unified bilateral-branch structure shown in Figure 3.	-
devise	vt. 设计；想出；发明；图谋；遗赠给	The solution is to devise transformation functions based on the intensity distribution of pixel neighborhoods.	design
seek to	追求；争取；力图	Two recent works, Ulyanov et al. (2016) Johnson et al. (2016), sought to address this problem by learning equivalent feed-forward generator networks that can generate the stylized image in a single pass.	-
discern	vi. 辨别 vt. 觉察出；识别；了解；隐约看见	Thus, by tracing the different behavior exhibited by natural images and manipulated samples, we can discern among them.	distinguish;
entail	vt. 使需要，必需；承担；遗传给；蕴含 n. 引起；需要；继承	Propensity score matching entails forming matched sets of treated and untreated subjects.	require
unveil	vt. 使公之于众，揭开；揭幕 vi. 除去面纱；显露	Figure 3 also unveils an important property for the distillation method.	uncover; disclose; reveal
reveal	vt. 显示；透露；揭露；泄露	Our study reveals that the optimal attack strategy depends on the energy budget and the singular values of the data matrix.	unveil; uncover; disclose
dub	v. 把...称为	We propose to leverage periodic activation functions for implicit neural representations and demonstrate that these networks, dubbed sinusoidal representation networks or SIRENs
amplify	vt. 放大，扩大；增强；详述 vi. 详述	we amplify the differences between the pairs with ...	enhance
alleviate	vt. 减轻，缓和	We use memcached to alleviate database load. BERT alleviates the previously mentioned unidirectionality constraint by using a "masked language model" ... Several variations have been proposed to alleviate this problem ...	mitigate; address
mitigate	vt. 使缓和，使减轻 vi. 减轻，缓和下来	Furthermore, many works have proposed robust PCA that can mitigate the impact of certain percentages of outliers and small dense random noise. To mitigate the problem of losing resolution for old memories, we introduce the concept of “sticky memories”	alleviate
ameliorate	vt. 改善；减轻（痛苦等）；改良	To ameliorate this pessimism, we propose a heuristic method.	mitigate
compromise	v. 妥协，折中；违背（原则），达不到（标准）； n. 妥协，和解；妥协（或折中）方案；达成妥协	The usage of BN often requires these systems to compromise between the model design and batch sizes.	trade-off
harvest	vt. 收割；得到 n. 收获；产量；结果	Those filters harvest information naturally from local neighborhoods in the input image	obtain
prohibit	vt. 阻止，禁止	This prohibits the unique partitioning of objects for all but the most specific purposes.	prohibitive
impede	vt. 阻碍；妨碍；阻止	This impedes the learning of those data-driven neural dialogue models.	hinder; prevent; hamper
confine	vt.限制, 局限于	We confine the related work to the domain of object recognition and divide it into three categories...	limit; restrict
restrict	vt. 限制；约束；限定	Most deep learning methods require substantial amounts of manually labeled data, which restricts their applicability in may domains that suffer from a dearth of annotated resources. Their usage of LSTM models restricts their prediction ability to a short range.	confine
abstain	v. （投票时）弃权，放弃；避免，回避；自制，戒除	These models are allowed to abstain whenever they are not sufficiently confident in their prediction. Thus, the selective model abstains from prediction at a point \(x\) iff \(g(x)=0\). The method of learning with a rejection option tries to avoid incorrect predictions by abstaining on some uncertain data.	abstain
mimic	vt. 模仿，摹拟 n. 善于模仿的人；会模拟的动物（或植物）adj. 模仿的，模拟的；假装的	that is, these models mimic the human conversations by training on the large-scale query-response pairs.	imitate
refer	vi. 参考；涉及；提到；查阅; vt. 涉及；委托；归诸于；使…求助于	If not further specified, we will use the term network, or the abbreviation NN, to refer to FC neural networks. Finally, we refer readers to [JNM20] and the references mentioned therein for an empirical study of ...	represent; denote

词组

phrase	含义	例句	近义词
reason about	推出;理出;思考得出	They reason about how best to defeat me.	infer; deduce
consist of	由…组成；由…构成；包括	The ROSA framework consists of a segment-wise shielding component, an FCN-based backbone network, and a context-aware restoration component.	be composed of
be composed of	由...组成	Objects are sometimes composed of heterogeneous colors and textures.	consist of
have access to	可以利用；使用；接近；	We don’t have access to the counterfactual outcomes.	-
resort to	依靠，求助于；诉诸；常去	One category is the conventional methods that do not use neural networks but resort to prior knowledge and handcrafted features	turn to
be capable of	能够; 有能力	In this paper, we propose a treatment effects estimation framework that is capable of adaptively exploiting observations from multiple sources to help infer causal effects in a target population with scarce data.	be able to
be prone to	易于……；有……的倾向	..., implying that using the cross-token MLP to learn the interplay across image patches is more prone to overfitting than ViTs' self-attention mechanism ...	tend
differ from	与...不同；区别于...	However, an observational study differs from a randomized experiment in one design issue.	-
differ in	不同在；在…方面存在不同	The different vegetational parts differ in their food qualities.	-
by its|their very nature	就其本质而言	By its very nature, beginners in digital image processing find enhancement to be interesting and relatively simple to understand.	by nature; in essence
for the sake of	为了；为了……的利益	For the sake of conceptual consistency, the objective function ...	in order to
with no/less regard for	不考虑...	Natural error alone can be minimized by using an optimal classifier with no regard for adversarial robustness and never abstaining.	with regard to; regarding; about; with respect to; in terms of; concerning
play a role in	扮演...的角色; 发挥...的作用	Unsupervised pre-training played a central role in the resurgence of deep learning.	-
on par with	与...同等水平	It is almost on par with RetinaNet and far above the SSD variants.	-

各种介词

at	-
Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step.
... the most adversarial data within the \(\epsilon\)-ball centered at \(x\) ...

in	-
Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks.
Deep neural networks trained with some version of Stochastic Gradient Descent have been shown to substantially outperform previous approaches on various supervised learning tasks in computer vision and speech processing

on	-
Deep neural networks trained with some version of Stochastic Gradient Descent have been shown to substantially outperform previous approaches on various supervised learning tasks in computer vision and speech processing

onto	-
The recent work of Gatys et al. (2016) introduced a method for transferring a style from an image onto another one, as demonstrated in fig. 1

with	-
These neighbors, together with the 4-neighbors, are called the 8-neighbors of p, denoted by \(N_8(p)\).

within	-
Unlike batch normalization, the proposed method directly estimates the normalization statistics from the summed inputs to the neurons within a hidden layer so the normalization does not introduce any new dependencies between training cases.

across	-
The normalization standardizes each summed input using its mean and its standard deviation across the training data.

over
It also has only one set of gain and bias parameters shared over all time-steps.
Then we optimize the expectation of the augmented loss parts over this distribution to approximate the original augmented loss parts computed with various augmented data mapping functions.
However, both of these approaches come at a non-trivial, additional computational cost, often increasing training time by an order of magnitude over standard training.

upon
Other mechanisms are built upon functions that serve as proxies for prediction confidence.	-

句子

摘要

sentence

Most deep learning methods require substantial amounts of manually labeled data, which restricts their applicability in may domains that suffer from a dearth of annotated resources.

Crafted adversarial data can easily fool the standard-trained deep models by adding human-imperceptible noise to the natural data, which leads to the security issue in applications such as medicine, finance, and autonomous driving.

Adversarial training (AT) is one of the most effective strategies for promoting model robustness, whereas even the state-of-the-art adversarially trained models struggle to exceed 60% robust test accuracy on CIFAR-10 without additional data, which is far from practical.

Detecting test samples drawn sufficiently far away from the training distribution statistically or adversarially is a fundamental requirement for deploying a good classifier in many real-world machine learning applications.

Qualitative and quantitative results are shown for both learning tasks, giving readers an insight on the importance of having AIM in the context of few-shot and continual learning.

We propose to leverage periodic activation functions for implicit neural representations and demonstrate that these networks, dubbed sinusoidal representation networks or SIRENs, are ideally suited for representing complex natural signals and their derivatives.

引言

sentence

There is a lack of consensus in the applied literature as to which variables to include in the propensity score model.

There is no general agreement among authors regarding where image processing stops and other related ares ...

There are no clear-cut boundaries in the continuum from image processing at one end to computer vision at the other.

Many existing defense baselines struggle to remove potential adversarial noises from input images, which is different from our proposed idea that adaptively predicts saliency maps for inputs with some new introduced generic noise to resist adversarial attacks.

There is a long history of pre-training general language representations, and we briefly review the most widely-used approaches in this section.

Batch Normalization (BN) is a milestone technique in the development of deep learning, enabling various networks to train.

in the last years, an exploding growth in studies of adversarial attacks and defenses has been witnessed.

This recent approach in object recognition ha made enormous progress in less than a decade.

Since [28], numerous works appeared to measure the saliency of pixels.

While the Transformer architecture has become the de-facto standard for natural language processing tasks, its applications to computer vision remain limited.

In response to this difficulty, there has been a recent surge in work that tries to to reduce the complexity of generating an adversarial example, which forms the bulk of the additional computation in adversarial training

However, many deep neural networks are vulnerable to various malicious perturbations due to their black-box nature: a small (even imperceptible) perturbation of input data may lead to completely wrong predictions

sentence

The remainder of the paper is organized as follows.

In this paper we revisit the method for feed-forward stylization of Ulyanov et al. (2016) and show that a small change in a generator architecture leads to much improved results.

Our high-level idea is to measure the probability density of test sample on feature spaces of DNNs ...

总结

sentence

Using error back-propagation method can be considered as a double-edged sword for deep neural networks.

正文

sentence

from a conceptual point of view; from a technical point of view

This is beyond the scope of the present paper.

recognizing those individual characters are in the scope of what we call digital image processing in this book ...

The non-negative L is a constant parameter that describes the tolerance level of varying age relationship.

... is a set C consisting of elements belonging either to A, to B, or to both.

The NOT of binary image B, is an array obtained by changing all 1-valued pixels to 0, and vice versa.

We use terms intensity and gray level interchangeably in subsequent discussions.

To better understand the nature of adversarial training and Theorem 1, we provide supporting schematics in Figure 2.

实验

sentence

When full-network fine-tuning on CIFAR-10 and CIFAR-100, we use the augmentation popularized by Wide Residual Networks: 4 pixels are reflection padded on each side, and a 32 x 32 crop is randomly sampled from the padded image or its horizontal flip.

We found that the exact value of the hyperparameter \(\alpha\) did not matter too much as long as it was roughly between 0.1 and 0.4.

We start with an initial ε value of 4/255 up to one-fourth the training epochs, and ramp up this value linearly to a value of 16/255 at the last epoch.

For example, as shown in the bottom panels of Figure 1

sentence

To compare different methods, we evaluate the minimum, average, and maximum l2 norm of the perturbation over all cases of wrong labels and report the average values over the test samples.

To avoid any leakage of the test set in generated samples

We adopt early-stopping strategy and terminate the training if the performance on validation set is not improved after 2 consecutive epochs.

The neural network is implemented with and enjoys efficiency from 1 × 1 convolutional layers.

As shown in Table I, Fβ -measure of DSS and DCL drop 30%–36% when exposed to the adversarial samples.

The adversarial attack reduces Fβ -measure of RFCN, Amulet, and UCF by 6%–12% while it only lowers Fβ of MC, LEGS, and MDF by 0.7%–1.1%.

Deep CNNs demonstrate a high fitting capacity superior to traditional methods using handcrafted features.

As Table II shows, DSS+ours outperforms DSS* by 71.9%–86.2% with respect to Fβ -measure on adversarial samples.

DSS+ours* still significantly surpasses DSS* by 71.86%–86.05% Fβ .

DSS+ours exceeds the second best DSS+TVM by 6.08% Fβ and 0.0337 MAE on the adversarial samples of HKU-IS dataset.

We find that as \(\beta\) increases, there is an increase in robustness, which plateaus at around \(\beta=0.3\).

To further justify the above fact, we plot output distributions of other layers

This signifies that it is increasingly difficult to find adversarial data that misclassifies the model.

This is somewhat expected since, on one hand, the robustness might not be affected much unless a sufficient number of filters (channels) are removed, which is different to depth that configures the entire residual block.

This can be explained by the gap between a pixel reconstruction task and a recognition task: the last several layers in an autoencoder are more specialized for reconstruction, but are less relevant for recognition.