Machine learning models have been found to learn shortcuts -- unintended decision rules that are unable to generalize -- undermining models' reliability. Previous works address this problem under the tenuous assumption that only a single shortcut exists in the training data. Real-world images are rife with multiple visual cues from background to texture. Key to advancing the reliability of vision systems is understanding whether existing methods can overcome multiple shortcuts or struggle in a Whac-A-Mole game, i.e., where mitigating one shortcut amplifies reliance on others. To address this shortcoming, we propose two benchmarks: 1) UrbanCars, a dataset with precisely controlled spurious cues, and 2) ImageNet-W, an evaluation set based on ImageNet for watermark, a shortcut we discovered affects nearly every modern vision model. Along with texture and background, ImageNet-W allows us to study multiple shortcuts emerging from training on natural images. We find computer vision models, including large foundation models -- regardless of training set, architecture, and supervision -- struggle when multiple shortcuts are present. Even methods explicitly designed to combat shortcuts struggle in a Whac-A-Mole dilemma. To tackle this challenge, we propose Last Layer Ensemble, a simple-yet-effective method to mitigate multiple shortcuts without Whac-A-Mole behavior. Our results surface multi-shortcut mitigation as an overlooked challenge critical to advancing the reliability of vision systems. The datasets and code are released: https://github.com/facebookresearch/Whac-A-Mole.git.

已经发现深层图像分类器可以从数据集中学习偏差。为了减轻偏见，大多数以前的方法都需要标签受保护的属性（例如，年龄，肤色）为全套，这有两个限制：1）当标签不可用时，它是不可行的； 2）它们无法缓解未知的偏见 - 人类没有先入为主的偏见。为了解决这些问题，我们提出了偏见的替代网络（Debian），该网络包括两个网络 - 一个发现者和一个分类器。通过以另一种方式培训，发现者试图找到分类器的多个未知偏见，而无需任何偏见注释，分类器的目的是删除发现者确定的偏见。虽然先前的作品评估了单个偏差的结果，但我们创建了多色MNIST数据集，以更好地缓解多偏差设置中的多个偏差，这不仅揭示了以前的方法中的问题，而且还展示了Debian的优势。在同时识别和减轻多种偏见时。我们进一步对现实世界数据集进行了广泛的实验，表明Debian中的发现者可以识别人类可能很难找到的未知偏见。关于辩护，Debian实现了强烈的偏见缓解绩效。

尽管最近通过剩余网络的代表学习中的自我监督方法取得了进展，但它们仍然对ImageNet分类基准进行了高度的监督学习，限制了它们在性能关键设置中的适用性。在MITROVIC等人的现有理论上洞察中建立2021年，我们提出了RELICV2，其结合了明确的不变性损失，在各种适当构造的数据视图上具有对比的目标。 Relicv2在ImageNet上实现了77.1％的前1个分类准确性，使用线性评估使用Reset50架构和80.6％，具有较大的Reset型号，优于宽边缘以前的最先进的自我监督方法。最值得注意的是，RelicV2是使用一系列标准Reset架构始终如一地始终优先于类似的对比较中的监督基线的第一个表示学习方法。最后，我们表明，尽管使用Reset编码器，Relicv2可与最先进的自我监控视觉变压器相媲美。

虽然大型审计的基础模型（FMS）对数据集级别的分布变化显示出显着的零击分类鲁棒性，但它们对亚群或组移动的稳健性相对却相对不受欢迎。我们研究了这个问题，并发现诸如剪辑之类的FMS可能对各种群体转移可能不健壮。在9个稳健性基准中，其嵌入式分类零射击分类导致平均和最差组精度之间的差距高达80.7个百分点（PP）。不幸的是，现有的改善鲁棒性的方法需要重新培训，这在大型基础模型上可能非常昂贵。我们还发现，改善模型推理的有效方法（例如，通过适配器，具有FM嵌入式作为输入的轻量级网络）不会持续改进，有时与零击相比会伤害组鲁棒性（例如，将精度差距提高到50.1 pp on 50.1 pp on On on 50.1 pp on Celeba）。因此，我们制定了一种适配器培训策略，以有效有效地改善FM组的鲁棒性。我们激励的观察是，尽管同一阶级中的群体中较差的鲁棒性在基础模型“嵌入空间”中分开，但标准适配器训练可能不会使这些要点更加紧密。因此，我们提出了对比度的适应，该适应器会通过对比度学习进行训练适配器，以使样品嵌入在同一类中的地面真相类嵌入和其他样品嵌入。在整个9个基准测试中，我们的方法始终提高组鲁棒性，使最差的组精度提高了8.5至56.0 pp。我们的方法也是有效的，这样做的方法也没有任何FM芬太尼，只有一组固定的冷冻FM嵌入。在水鸟和Celeba等基准上，这导致最差的组精度可与最先进的方法相媲美，而最先进的方法可以重新训练整个模型，而仅训练$ \ leq $ 1％的模型参数。

Contrastive Language-Image Pre-trained (CLIP) models have zero-shot ability of classifying an image belonging to "[CLASS]" by using similarity between the image and the prompt sentence "a [CONTEXT] of [CLASS]". Based on exhaustive text cues in "[CONTEXT]", CLIP model is aware of different contexts, e.g. background, style, viewpoint, and exhibits unprecedented robustness against a wide range of distribution shifts. However, recent works find further fine-tuning of CLIP models improves accuracy but sacrifices the robustness on downstream tasks. We conduct an empirical investigation to show fine-tuning will corrupt the context-aware ability of pre-trained CLIP features. To solve this problem, we propose Context-Aware Robust Fine-tuning (CAR-FT). CAR-FT regularizes the model during fine-tuning to capture the context information. Specifically, we use zero-shot prompt weights to get the context distribution contained in the image. By minimizing the Kullback-Leibler Divergence (KLD) between context distributions induced by original/fine-tuned CLIP models, CAR-FT makes the context-aware ability of CLIP inherited into downstream tasks, and achieves both higher In-Distribution (ID) and Out-Of-Distribution (OOD) accuracy. The experimental results show CAR-FT achieves superior robustness on five OOD test datasets of ImageNet, and meanwhile brings accuracy gains on nine downstream tasks. Additionally, CAR-FT surpasses previous Domain Generalization (DG) methods and gets 78.5% averaged accuracy on DomainBed benchmark, building the new state-of-the-art.

最近证明，接受SGD训练的神经网络优先依赖线性预测的特征，并且可以忽略复杂的，同样可预测的功能。这种简单性偏见可以解释他们缺乏分布（OOD）的鲁棒性。学习任务越复杂，统计工件（即选择偏见，虚假相关性）的可能性就越大比学习的机制更简单。我们证明可以减轻简单性偏差并改善了OOD的概括。我们使用对其输入梯度对齐的惩罚来训练一组类似的模型以不同的方式拟合数据。我们从理论和经验上展示了这会导致学习更复杂的预测模式的学习。 OOD的概括从根本上需要超出I.I.D.示例，例如多个培训环境，反事实示例或其他侧面信息。我们的方法表明，我们可以将此要求推迟到独立的模型选择阶段。我们获得了SOTA的结果，可以在视觉域偏置数据和概括方面进行视觉识别。该方法 - 第一个逃避简单性偏见的方法 - 突出了需要更好地理解和控制深度学习中的归纳偏见。

分发班次的稳健性对于部署现实世界中的机器学习模型至关重要。尽管如此必要的，但在定义导致这些变化的潜在机制以及评估跨多个不同的分发班次的稳健性的潜在机制很少。为此，我们介绍了一种框架，可实现各种分布换档的细粒度分析。我们通过评估在合成和现实世界数据集中分为五个类别的19个不同的方法来提供对当前最先进的方法的整体分析。总的来说，我们训练超过85架模型。我们的实验框架可以很容易地扩展到包括新方法，班次和数据集。我们发现，与以前的工作〜\ citep {gulrajani20}不同，该进度已经通过标准的ERM基线进行;特别是，在许多情况下，预先训练和增强（学习或启发式）提供了大的收益。但是，最好的方法在不同的数据集和班次上不一致。

我们有兴趣从数据不足的情况下学习强大的模型，而无需任何外部预训练的检查点。首先，与足够的数据相比，我们展示了为什么数据不足会使模型更容易偏向于通常不同于测试的有限培训环境。例如，如果所有训练天鹅样本都是“白色”，则该模型可能错误地使用“白色”环境来代表内在的天鹅。然后，我们证明，均衡感应偏差可以保留类功能，而不变性电感偏差可以消除环境功能，从而使类功能概括为测试中的任何环境变化。为了将它们强加于学习，我们证明可以部署任何基于现成的基于对比度的自我监督特征学习方法；对于不变性，我们提出了一个范围的不变风险最小化（IRM），该风险最小化（IRM）有效地应对常规IRM中缺少环境注释的挑战。对现实世界基准（Vipriors，Imagenet100和Nico）的最新实验结果验证了在数据效率学习中的巨大潜力和不变性的潜力。该代码可从https://github.com/wangt-cn/eqinv获得

We present a framework for ranking images within their class based on the strength of spurious cues present. By measuring the gap in accuracy on the highest and lowest ranked images (we call this spurious gap), we assess spurious feature reliance for $89$ diverse ImageNet models, finding that even the best models underperform in images with weak spurious presence. However, the effect of spurious cues varies far more dramatically across classes, emphasizing the crucial, often overlooked, class-dependence of the spurious correlation problem. While most spurious features we observe are clarifying (i.e. improving test-time accuracy when present, as is typically expected), we surprisingly find many cases of confusing spurious features, where models perform better when they are absent. We then close the spurious gap by training new classification heads on lowly ranked (i.e. without common spurious cues) images, resulting in improved effective robustness to distribution shifts (ObjectNet, ImageNet-R, ImageNet-Sketch). We also propose a second metric to assess feature reliability, finding that spurious features are generally less reliable than non-spurious (core) ones, though again, spurious features can be more reliable for certain classes. To enable our analysis, we annotated $5,000$ feature-class dependencies over {\it all} of ImageNet as core or spurious using minimal human supervision. Finally, we show the feature discovery and spuriosity ranking framework can be extended to other datasets like CelebA and WaterBirds in a lightweight fashion with only linear layer training, leading to discovering a previously unknown racial bias in the Celeb-A hair classification.

自我监督学习的最新进展证明了多种视觉任务的有希望的结果。高性能自我监督方法中的一个重要成分是通过培训模型使用数据增强，以便在嵌入空间附近的相同图像的不同增强视图。然而，常用的增强管道整体地对待图像，忽略图像的部分的语义相关性-e.g。主题与背景 - 这可能导致学习杂散相关性。我们的工作通过调查一类简单但高度有效的“背景增强”来解决这个问题，这鼓励模型专注于语义相关内容，劝阻它们专注于图像背景。通过系统的调查，我们表明背景增强导致在各种任务中跨越一系列最先进的自我监督方法（MOCO-V2，BYOL，SWAV）的性能大量改进。 $ \ SIM $ + 1-2％的ImageNet收益，使得与监督基准的表现有关。此外，我们发现有限标签设置的改进甚至更大（高达4.2％）。背景技术增强还改善了许多分布换档的鲁棒性，包括天然对抗性实例，想象群-9，对抗性攻击，想象成型。我们还在产生了用于背景增强的显着掩模的过程中完全无监督的显着性检测进展。

The well-documented presence of texture bias in modern convolutional neural networks has led to a plethora of algorithms that promote an emphasis on shape cues, often to support generalization to new domains. Yet, common datasets, benchmarks and general model selection strategies are missing, and there is no agreed, rigorous evaluation protocol. In this paper, we investigate difficulties and limitations when training networks with reduced texture bias. In particular, we also show that proper evaluation and meaningful comparisons between methods are not trivial. We introduce BiasBed, a testbed for texture- and style-biased training, including multiple datasets and a range of existing algorithms. It comes with an extensive evaluation protocol that includes rigorous hypothesis testing to gauge the significance of the results, despite the considerable training instability of some style bias methods. Our extensive experiments, shed new light on the need for careful, statistically founded evaluation protocols for style bias (and beyond). E.g., we find that some algorithms proposed in the literature do not significantly mitigate the impact of style bias at all. With the release of BiasBed, we hope to foster a common understanding of consistent and meaningful comparisons, and consequently faster progress towards learning methods free of texture bias. Code is available at https://github.com/D1noFuzi/BiasBed

这项工作的目的是在训练过程中划分和名称图像区域，而无需访问像素级标签。为了解决这项任务，我们通过提炼两个基础模型的互补优势来构建细分器。第一个剪辑（Radford等，2021）具有将名称分配给图像内容的能力，但缺乏对象结构的可访问表示。第二个Dino（Caron等，2021）捕获了物体的空间范围，但对对象名称不了解。我们的方法称为名为Mask，开始使用剪辑来构建特定于类别的图像档案。这些图像用dino的类别 - 敏捷的对象检测器进行伪标记，然后使用夹档案标签通过类别特定的细分器进行完善。得益于精制面具的高质量，我们表明，在这些档案中训练有适当数据的培训的标准分割体系结构可为单对象和多对象图像带来令人印象深刻的语义细分能力。结果，我们提出的名字命名为在包括VOC2012，可可和大规模ImageNet-S数据集在内的五个基准上的一系列先前工作中表现出色。

State-of-the-art computer vision systems are trained to predict a fixed set of predetermined object categories. This restricted form of supervision limits their generality and usability since additional labeled data is needed to specify any other visual concept. Learning directly from raw text about images is a promising alternative which leverages a much broader source of supervision. We demonstrate that the simple pre-training task of predicting which caption goes with which image is an efficient and scalable way to learn SOTA image representations from scratch on a dataset of 400 million (image, text) pairs collected from the internet. After pre-training, natural language is used to reference learned visual concepts (or describe new ones) enabling zero-shot transfer of the model to downstream tasks. We study the performance of this approach by benchmarking on over 30 different existing computer vision datasets, spanning tasks such as OCR, action recognition in videos, geo-localization, and many types of fine-grained object classification. The model transfers non-trivially to most tasks and is often competitive with a fully supervised baseline without the need for any dataset specific training. For instance, we match the accuracy of the original ResNet-50 on ImageNet zero-shot without needing to use any of the 1.28 million training examples it was trained on. We release our code and pre-trained model weights at https://github.com/OpenAI/CLIP.

对比训练有素的语言图像模型，例如剪辑，Align和Basic，已经证明了对多种具有挑战性的自然分配变化的前所未有的鲁棒性。由于这些语言图像模型与以前的培训方法有多种不同，因此一个重要的问题是导致稳定性增长的原因。我们通过系统的实验研究回答这个问题。具体而言，我们研究了鲁棒性增长的五个不同可能的原因：（i）训练集大小，（ii）培训分配，（iii）在培训时进行语言监督，（iv）测试时语言监督，以及（v）对比损失函数。我们的实验表明，更多样化的训练分布是稳健性增长的主要原因，其他因素几乎没有稳健性。除了实验结果之外，我们还引入了Imagenet捕获，这是一种来自Flickr的原始文本注释的Imagenet版本，以实现语言图像训练的进一步受控实验。

预训练的视觉模型（例如，剪辑）在许多下游任务中显示出有希望的零弹性概括，并具有正确设计的文本提示。最近的作品不依赖手工设计的提示，而是使用下游任务的培训数据来学习提示。虽然有效，但针对领域数据的培训却降低了模型的概括能力，使其无法看到新领域。在这项工作中，我们提出了测试时间提示调整（TPT），该方法可以通过单个测试样本即时学习自适应提示。对于图像分类，TPT通过使用置信度选择最小化熵来优化提示，以便模型在每个测试样本的不同增强视图上都具有一致的预测。在评估对自然分布变化的概括时，TPT平均将零击的TOP-1精度提高了3.6％，超过了先前需要其他特定于任务的训练数据的迅速调整方法。在评估看不见类别的跨数据集泛化时，TPT与使用其他培训数据的最先进方法相当。项目页面：https：//azshue.github.io/tpt。

我们介绍了几个新的数据集即想象的A / O和Imagenet-R以及合成环境和测试套件，我们称为CAOS。 Imagenet-A / O允许研究人员专注于想象成剩余的盲点。由于追踪稳健的表示，以特殊创建了ImageNet-R，因为表示不再简单地自然，而是包括艺术和其他演绎。 Caos Suite由Carla Simulator构建，允许包含异常物体，可以创建可重复的合成环境和用于测试稳健性的场景。所有数据集都是为测试鲁棒性和衡量鲁棒性的衡量进展而创建的。数据集已用于各种其他作品中，以衡量其具有鲁棒性的自身进步，并允许切向进展，这些进展不会完全关注自然准确性。鉴于这些数据集，我们创建了几种旨在推进鲁棒性研究的新方法。我们以最大Logit的形式和典型程度的形式构建简单的基线，并以深度的形式创建新的数据增强方法，从而提高上述基准。最大Logit考虑Logit值而不是SoftMax操作后的值，而微小的变化会产生明显的改进。典型程分将输出分布与类的后部分布进行比较。我们表明，除了分段任务之外，这将提高对基线的性能。猜测可能在像素级别，像素的语义信息比类级信息的语义信息不太有意义。最后，新的Deepaulment的新增强技术利用神经网络在彻底不同于先前使用的传统几何和相机的转换的图像上创建增强。

分布式概括（OOD）都是关于对环境变化的学习不变性。如果每个类中的上下文分布均匀分布，则OOD将是微不足道的，因为由于基本原则，可以轻松地删除上下文：类是上下文不变的。但是，收集这种平衡的数据集是不切实际的。学习不平衡的数据使模型偏见对上下文，从而伤害了OOD。因此，OOD的关键是上下文平衡。我们认为，在先前工作中广泛采用的假设，可以直接从偏见的类预测中注释或估算上下文偏差，从而使上下文不完整甚至不正确。相比之下，我们指出了上述原则的另一面：上下文对于类也不变，这激励我们将类（已经被标记为已标记的）视为不同环境以解决上下文偏见（没有上下文标签）。我们通过最大程度地减少阶级样本相似性的对比损失，同时确保这种相似性在所有类别中不变，从而实现这一想法。在具有各种上下文偏见和域间隙的基准测试中，我们表明，配备了我们上下文估计的简单基于重新加权的分类器实现了最新的性能。我们在https://github.com/simpleshinobu/irmcon上提供了附录中的理论理由和代码。

We introduce four new real-world distribution shift datasets consisting of changes in image style, image blurriness, geographic location, camera operation, and more. With our new datasets, we take stock of previously proposed methods for improving out-of-distribution robustness and put them to the test. We find that using larger models and artificial data augmentations can improve robustness on realworld distribution shifts, contrary to claims in prior work. We find improvements in artificial robustness benchmarks can transfer to real-world distribution shifts, contrary to claims in prior work. Motivated by our observation that data augmentations can help with real-world distribution shifts, we also introduce a new data augmentation method which advances the state-of-the-art and outperforms models pretrained with 1000× more labeled data. Overall we find that some methods consistently help with distribution shifts in texture and local image statistics, but these methods do not help with some other distribution shifts like geographic changes. Our results show that future research must study multiple distribution shifts simultaneously, as we demonstrate that no evaluated method consistently improves robustness.

Models trained via empirical risk minimization (ERM) are known to rely on spurious correlations between labels and task-independent input features, resulting in poor generalization to distributional shifts. Group distributionally robust optimization (G-DRO) can alleviate this problem by minimizing the worst-case loss over a set of pre-defined groups over training data. G-DRO successfully improves performance of the worst-group, where the correlation does not hold. However, G-DRO assumes that the spurious correlations and associated worst groups are known in advance, making it challenging to apply it to new tasks with potentially multiple unknown spurious correlations. We propose AGRO -- Adversarial Group discovery for Distributionally Robust Optimization -- an end-to-end approach that jointly identifies error-prone groups and improves accuracy on them. AGRO equips G-DRO with an adversarial slicing model to find a group assignment for training examples which maximizes worst-case loss over the discovered groups. On the WILDS benchmark, AGRO results in 8% higher model performance on average on known worst-groups, compared to prior group discovery approaches used with G-DRO. AGRO also improves out-of-distribution performance on SST2, QQP, and MS-COCO -- datasets where potential spurious correlations are as yet uncharacterized. Human evaluation of ARGO groups shows that they contain well-defined, yet previously unstudied spurious correlations that lead to model errors.

Does the dominant approach to learn representations (as a side effect of optimizing an expected cost for a single training distribution) remain a good approach when we are dealing with multiple distributions. Our thesis is that such scenarios are better served by representations that are "richer" than those obtained with a single optimization episode. This is supported by a collection of empirical results obtained with an apparently na\"ive ensembling technique: concatenating the representations obtained with multiple training episodes using the same data, model, algorithm, and hyper-parameters, but different random seeds. These independently trained networks perform similarly. Yet, in a number of scenarios involving new distributions, the concatenated representation performs substantially better than an equivalently sized network trained from scratch. This proves that the representations constructed by multiple training episodes are in fact different. Although their concatenation carries little additional information about the training task under the training distribution, it becomes substantially more informative when tasks or distributions change. Meanwhile, a single training episode is unlikely to yield such a redundant representation because the optimization process has no reason to accumulate features that do not incrementally improve the training performance.