我们在变压器中重新审视设计选择，并提出方法来解决它们在处理长序列中的弱点。首先，我们提出了一个名为“门控注意单元”的简单层，该层允许使用较弱的单头注意，而质量损失最小。然后，我们提出了一种与该新层的线性近似方法互补的，该方法对加速器友好且质量高度竞争。最终的型号（名为Flash）与短（512）和长（8K）上下文长度相匹配，在WIKI-40B上达到高达4.9 $ \ times $的训练速度和PG上的12.1 $ \ times $，在PG上达到了4.9 $ \ times $的困惑。-19用于自动回归语言建模，C4的4.8 $ \ times $用于掩盖语言建模。

我们提出了一种称为基本的组合缩放方法，可在ImageNet ILSVRC-2012验证集上实现85.7％的前1个零点精度，超越了最佳发布的零拍模型 - 剪辑并对齐 - 达9.3％。我们的基本模式还显示出鲁棒性基准的显着改进。例如，在5个测试集中，具有自然分布换档，如想象的 - {A，R，V2，素描}和ObjectNet，我们的车型实现了83.7％的前1个平均精度，只有一个小幅度从其原始的想象精度下降。为实现这些结果，我们扩大了剪辑的对比学习框架，并在三个方面对齐：数据大小，型号大小和批量大小。我们的数据集具有6.6B噪声图像文本对，比对齐的4倍，比夹子大16倍。我们最大的型号具有3B重量，参数比为3.75倍，拖鞋比对齐和夹子更大。我们的批量尺寸为65536，比剪辑的2倍，4倍超过对齐。缩放的主要挑战是我们的加速器的内存有限，如GPU和TPU。因此，我们提出了一种在线渐变缓存的简单方法来克服这个限制。

Transformers have attracted increasing interests in computer vision, but they still fall behind state-of-the-art convolutional networks. In this work, we show that while Transformers tend to have larger model capacity, their generalization can be worse than convolutional networks due to the lack of the right inductive bias. To effectively combine the strengths from both architectures, we present CoAtNets (pronounced "coat" nets), a family of hybrid models built from two key insights:(1) depthwise Convolution and self-Attention can be naturally unified via simple relative attention; (2) vertically stacking convolution layers and attention layers in a principled way is surprisingly effective in improving generalization, capacity and efficiency. Experiments show that our CoAtNets achieve state-of-the-art performance under different resource constraints across various datasets: Without extra data, CoAtNet achieves 86.0% ImageNet top-1 accuracy; When pre-trained with 13M images from ImageNet-21K, our CoAtNet achieves 88.56% top-1 accuracy, matching ViT-huge pre-trained with 300M images from JFT-300M while using 23x less data; Notably, when we further scale up CoAtNet with JFT-3B, it achieves 90.88% top-1 accuracy on ImageNet, establishing a new state-of-the-art result.1 The initial projection stage can be seen as an aggressive down-sampling convolutional stem.

With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves better performance than pretraining approaches based on autoregressive language modeling. However, relying on corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model, into pretraining. Empirically, under comparable experiment settings, XLNet outperforms BERT on 20 tasks, often by a large margin, including question answering, natural language inference, sentiment analysis, and document ranking. 1 .

Semi-supervised learning lately has shown much promise in improving deep learning models when labeled data is scarce. Common among recent approaches is the use of consistency training on a large amount of unlabeled data to constrain model predictions to be invariant to input noise. In this work, we present a new perspective on how to effectively noise unlabeled examples and argue that the quality of noising, specifically those produced by advanced data augmentation methods, plays a crucial role in semi-supervised learning. By substituting simple noising operations with advanced data augmentation methods such as RandAugment and back-translation, our method brings substantial improvements across six language and three vision tasks under the same consistency training framework. On the IMDb text classification dataset, with only 20 labeled examples, our method achieves an error rate of 4.20, outperforming the state-of-the-art model trained on 25,000 labeled examples. On a standard semi-supervised learning benchmark, CIFAR-10, our method outperforms all previous approaches and achieves an error rate of 5.43 with only 250 examples. Our method also combines well with transfer learning, e.g., when finetuning from BERT, and yields improvements in high-data regime, such as ImageNet, whether when there is only 10% labeled data or when a full labeled set with 1.3M extra unlabeled examples is used. 1

When using LiDAR semantic segmentation models for safety-critical applications such as autonomous driving, it is essential to understand and improve their robustness with respect to a large range of LiDAR corruptions. In this paper, we aim to comprehensively analyze the robustness of LiDAR semantic segmentation models under various corruptions. To rigorously evaluate the robustness and generalizability of current approaches, we propose a new benchmark called SemanticKITTI-C, which features 16 out-of-domain LiDAR corruptions in three groups, namely adverse weather, measurement noise and cross-device discrepancy. Then, we systematically investigate 11 LiDAR semantic segmentation models, especially spanning different input representations (e.g., point clouds, voxels, projected images, and etc.), network architectures and training schemes. Through this study, we obtain two insights: 1) We find out that the input representation plays a crucial role in robustness. Specifically, under specific corruptions, different representations perform variously. 2) Although state-of-the-art methods on LiDAR semantic segmentation achieve promising results on clean data, they are less robust when dealing with noisy data. Finally, based on the above observations, we design a robust LiDAR segmentation model (RLSeg) which greatly boosts the robustness with simple but effective modifications. It is promising that our benchmark, comprehensive analysis, and observations can boost future research in robust LiDAR semantic segmentation for safety-critical applications.

In recent years, arbitrary image style transfer has attracted more and more attention. Given a pair of content and style images, a stylized one is hoped that retains the content from the former while catching style patterns from the latter. However, it is difficult to simultaneously keep well the trade-off between the content details and the style features. To stylize the image with sufficient style patterns, the content details may be damaged and sometimes the objects of images can not be distinguished clearly. For this reason, we present a new transformer-based method named STT for image style transfer and an edge loss which can enhance the content details apparently to avoid generating blurred results for excessive rendering on style features. Qualitative and quantitative experiments demonstrate that STT achieves comparable performance to state-of-the-art image style transfer methods while alleviating the content leak problem.

With the increasing ability of large language models (LLMs), in-context learning (ICL) has become a new paradigm for natural language processing (NLP), where LLMs make predictions only based on contexts augmented with a few training examples. It has been a new trend exploring ICL to evaluate and extrapolate the ability of LLMs. In this paper, we aim to survey and summarize the progress, challenges, and future work in ICL. We first present a formal definition of ICL and clarify its correlation to related studies. Then, we organize and discuss advanced techniques of ICL, including training strategies, prompting strategies, and so on. Finally, we present the challenges of ICL and provide potential directions for further research. We hope our work can encourage more research on uncovering how ICL works and improving ICL in future work.

Gaze estimation is the fundamental basis for many visual tasks. Yet, the high cost of acquiring gaze datasets with 3D annotations hinders the optimization and application of gaze estimation models. In this work, we propose a novel Head-Eye redirection parametric model based on Neural Radiance Field, which allows dense gaze data generation with view consistency and accurate gaze direction. Moreover, our head-eye redirection parametric model can decouple the face and eyes for separate neural rendering, so it can achieve the purpose of separately controlling the attributes of the face, identity, illumination, and eye gaze direction. Thus diverse 3D-aware gaze datasets could be obtained by manipulating the latent code belonging to different face attributions in an unsupervised manner. Extensive experiments on several benchmarks demonstrate the effectiveness of our method in domain generalization and domain adaptation for gaze estimation tasks.

Generalizability to unseen forgery types is crucial for face forgery detectors. Recent works have made significant progress in terms of generalization by synthetic forgery data augmentation. In this work, we explore another path for improving the generalization. Our goal is to reduce the features that are easy to learn in the training phase, so as to reduce the risk of overfitting on specific forgery types. Specifically, in our method, a teacher network takes as input the face images and generates an attention map of the deep features by a diverse multihead attention ViT. The attention map is used to guide a student network to focus on the low-attended features by reducing the highly-attended deep features. A deep feature mixup strategy is also proposed to synthesize forgeries in the feature domain. Experiments demonstrate that, without data augmentation, our method is able to achieve promising performances on unseen forgeries and highly compressed data.