变化自动编码器（VAE）的最新进展使学习潜流歧管成为紧凑的谎言组，例如$ SO（d）$。由于这种方法假定数据在于谎言组本身同构的子空间，因此我们在这里研究了该假设如何在图像的背景下通过预测$ d $二维量产生的图像，而$ d $ d $ d $二维构成$ so so so so（d）$。在检查小组和图像空间的不同理论候选者后，我们表明，定义对数据空间的组动作的尝试通常会失败，因为它需要对卷上的更具体的几何约束。使用几何VAE，我们的实验证实了此约束是适当姿势推断的关键，我们讨论了这些结果对应用和未来工作的潜力。

State space models (SSMs) have demonstrated state-of-the-art sequence modeling performance in some modalities, but underperform attention in language modeling. Moreover, despite scaling nearly linearly in sequence length instead of quadratically, SSMs are still slower than Transformers due to poor hardware utilization. In this paper, we make progress on understanding the expressivity gap between SSMs and attention in language modeling, and on reducing the hardware barrier between SSMs and attention. First, we use synthetic language modeling tasks to understand the gap between SSMs and attention. We find that existing SSMs struggle with two capabilities: recalling earlier tokens in the sequence and comparing tokens across the sequence. To understand the impact on language modeling, we propose a new SSM layer, H3, that is explicitly designed for these abilities. H3 matches attention on the synthetic languages and comes within 0.4 PPL of Transformers on OpenWebText. Furthermore, a hybrid 125M-parameter H3-attention model that retains two attention layers surprisingly outperforms Transformers on OpenWebText by 1.0 PPL. Next, to improve the efficiency of training SSMs on modern hardware, we propose FlashConv. FlashConv uses a fused block FFT algorithm to improve efficiency on sequences up to 8K, and introduces a novel state passing algorithm that exploits the recurrent properties of SSMs to scale to longer sequences. FlashConv yields 2$\times$ speedup on the long-range arena benchmark and allows hybrid language models to generate text 1.6$\times$ faster than Transformers. Using FlashConv, we scale hybrid H3-attention language models up to 1.3B parameters on the Pile and find promising initial results, achieving lower perplexity than Transformers and outperforming Transformers in zero- and few-shot learning on a majority of tasks in the SuperGLUE benchmark.

Out-of-distribution (OOD) generalisation aims to build a model that can well generalise its learnt knowledge from source domains to an unseen target domain. However, current image classification models often perform poorly in the OOD setting due to statistically spurious correlations learning from model training. From causality-based perspective, we formulate the data generation process in OOD image classification using a causal graph. On this graph, we show that prediction P(Y|X) of a label Y given an image X in statistical learning is formed by both causal effect P(Y|do(X)) and spurious effects caused by confounding features (e.g., background). Since the spurious features are domain-variant, the prediction P(Y|X) becomes unstable on unseen domains. In this paper, we propose to mitigate the spurious effect of confounders using front-door adjustment. In our method, the mediator variable is hypothesized as semantic features that are essential to determine a label for an image. Inspired by capability of style transfer in image generation, we interpret the combination of the mediator variable with different generated images in the front-door formula and propose novel algorithms to estimate it. Extensive experimental results on widely used benchmark datasets verify the effectiveness of our method.

Diffusion models are rising as a powerful solution for high-fidelity image generation, which exceeds GANs in quality in many circumstances. However, their slow training and inference speed is a huge bottleneck, blocking them from being used in real-time applications. A recent DiffusionGAN method significantly decreases the models' running time by reducing the number of sampling steps from thousands to several, but their speeds still largely lag behind the GAN counterparts. This paper aims to reduce the speed gap by proposing a novel wavelet-based diffusion structure. We extract low-and-high frequency components from both image and feature levels via wavelet decomposition and adaptively handle these components for faster processing while maintaining good generation quality. Furthermore, we propose to use a reconstruction term, which effectively boosts the model training convergence. Experimental results on CelebA-HQ, CIFAR-10, LSUN-Church, and STL-10 datasets prove our solution is a stepping-stone to offering real-time and high-fidelity diffusion models. Our code and pre-trained checkpoints will be available at \url{https://github.com/VinAIResearch/WaveDiff.git}.

End-to-end multilingual ASR has become more appealing because of several reasons such as simplifying the training and deployment process and positive performance transfer from high-resource to low-resource languages. However, scaling up the number of languages, total hours, and number of unique tokens is not a trivial task. This paper explores large-scale multilingual ASR models on 70 languages. We inspect two architectures: (1) Shared embedding and output and (2) Multiple embedding and output model. In the shared model experiments, we show the importance of tokenization strategy across different languages. Later, we use our optimal tokenization strategy to train multiple embedding and output model to further improve our result. Our multilingual ASR achieves 13.9%-15.6% average WER relative improvement compared to monolingual models. We show that our multilingual ASR generalizes well on an unseen dataset and domain, achieving 9.5% and 7.5% WER on Multilingual Librispeech (MLS) with zero-shot and finetuning, respectively.

Scene Graph Generation (SGG) serves a comprehensive representation of the images for human understanding as well as visual understanding tasks. Due to the long tail bias problem of the object and predicate labels in the available annotated data, the scene graph generated from current methodologies can be biased toward common, non-informative relationship labels. Relationship can sometimes be non-mutually exclusive, which can be described from multiple perspectives like geometrical relationships or semantic relationships, making it even more challenging to predict the most suitable relationship label. In this work, we proposed the SG-Shuffle pipeline for scene graph generation with 3 components: 1) Parallel Transformer Encoder, which learns to predict object relationships in a more exclusive manner by grouping relationship labels into groups of similar purpose; 2) Shuffle Transformer, which learns to select the final relationship labels from the category-specific feature generated in the previous step; and 3) Weighted CE loss, used to alleviate the training bias caused by the imbalanced dataset.

Robots have been brought to work close to humans in many scenarios. For coexistence and collaboration, robots should be safe and pleasant for humans to interact with. To this end, the robots could be both physically soft with multimodal sensing/perception, so that the robots could have better awareness of the surrounding environment, as well as to respond properly to humans' action/intention. This paper introduces a novel soft robotic link, named ProTac, that possesses multiple sensing modes: tactile and proximity sensing, based on computer vision and a functional material. These modalities come from a layered structure of a soft transparent silicon skin, a polymer dispersed liquid crystal (PDLC) film, and reflective markers. Here, the PDLC film can switch actively between the opaque and the transparent state, from which the tactile sensing and proximity sensing can be obtained by using cameras solely built inside the ProTac link. In this paper, inference algorithms for tactile proximity perception are introduced. Evaluation results of two sensing modalities demonstrated that, with a simple activation strategy, ProTac link could effectively perceive useful information from both approaching and in-contact obstacles. The proposed sensing device is expected to bring in ultimate solutions for design of robots with softness, whole-body and multimodal sensing, and safety control strategies.

语义分割是开发医学图像诊断系统的重要任务。但是，构建注释的医疗数据集很昂贵。因此，在这种情况下，半监督方法很重要。在半监督学习中，标签的质量在模型性能中起着至关重要的作用。在这项工作中，我们提出了一种新的伪标签策略，可提高用于培训学生网络的伪标签的质量。我们遵循多阶段的半监督训练方法，该方法在标记的数据集上训练教师模型，然后使用训练有素的老师将伪标签渲染用于学生培训。通过这样做，伪标签将被更新，并且随着培训的进度更加精确。上一个和我们的方法之间的关键区别在于，我们在学生培训过程中更新教师模型。因此，在学生培训过程中，提高了伪标签的质量。我们还提出了一种简单但有效的策略，以使用动量模型来提高伪标签的质量 - 训练过程中原始模型的慢复制版本。通过应用动量模型与学生培训期间的重新渲染伪标签相结合，我们在五个数据集中平均达到了84.1％的骰子分数（即Kvarsir，CVC-ClinicdB，Etis-laribpolypdb，cvc-colondb，cvc-colondb，cvc-colondb和cvc-300）和CVC-300）只有20％的数据集用作标记数据。我们的结果超过了3％的共同实践，甚至在某些数据集中取得了完全监督的结果。我们的源代码和预培训模型可在https://github.com/sun-asterisk-research/online学习SSL上找到

使用通过组成可逆层获得的地图进行标准化模型复杂概率分布。特殊的线性层（例如蒙版和1x1卷积）在现有体系结构中起着关键作用，因为它们在具有可拖动的Jacobians和倒置的同时增加表达能力。我们提出了一个基于蝴蝶层的新的可逆线性层家族，理论上捕获复杂的线性结构，包括排列和周期性，但可以有效地倒置。这种代表力是我们方法的关键优势，因为这些结构在许多现实世界数据集中很常见。根据我们的可逆蝴蝶层，我们构建了一个新的称为蝴蝶流的归一化流量模型。从经验上讲，我们证明蝴蝶不仅可以在MNIST，CIFAR-10和Imagenet 32​​x32等自然图像上实现强密度估计结果，而且还可以在结构化数据集中获得明显更好的对数可能性，例如Galaxy图像和Mimic-III患者群体 - - 同时，在记忆和计算方面比相关基线更有效。

人类利用先验知识来描述图像，并能够使其解释适应特定的上下文信息，即使在上下文信息和图像不匹配时，也可以在发明合理的解释的范围内。在这项工作中，我们提出了通过整合上下文知识来字幕Wikipedia图像的新颖任务。具体而言，我们制作的模型共同推理了Wikipedia文章，Wikimedia图像及其相关描述以产生上下文化的标题。特别是，可以使用类似的Wikimedia图像来说明不同的文章，并且所产生的标题需要适应特定的上下文，因此使我们能够探索模型的限制以调整标题为不同的上下文信息。该领域中的一个特殊挑战性的任务是处理量不多的单词和命名实体。为了解决这个问题，我们提出了一个预训练目标，掩盖了命名实体建模（MNEM），并表明与基线模型相比，此借口任务可以改善。此外，我们验证了Wikipedia中使用MNEM目标预先训练的模型可以很好地推广到新闻字幕数据集。此外，我们根据字幕任务的难度定义了两种不同的测试拆分。我们提供有关每种方式的作用和重要性的见解，并突出我们模型的局限性。接受时，代码，模型和数据拆分可公开可用。