We introduce LaViLa, a new approach to learning video-language representations by leveraging Large Language Models (LLMs). We repurpose pre-trained LLMs to be conditioned on visual input, and finetune them to create automatic video narrators. Our auto-generated narrations offer a number of advantages, including dense coverage of long videos, better temporal synchronization of the visual information and text, and much higher diversity of text. The video-text embedding learned contrastively with these additional auto-generated narrations outperforms the previous state-of-the-art on multiple first-person and third-person video tasks, both in zero-shot and finetuned setups. Most notably, LaViLa obtains an absolute gain of 10.1% on EGTEA classification and 5.9% Epic-Kitchens-100 multi-instance retrieval benchmarks. Furthermore, LaViLa trained with only half the narrations from the Ego4D dataset outperforms baseline models trained on the full set, and shows positive scaling behavior on increasing pre-training data and model size.

Damage to the inferior frontal gyrus (Broca's area) can cause agrammatic aphasia wherein patients, although able to comprehend, lack the ability to form complete sentences. This inability leads to communication gaps which cause difficulties in their daily lives. The usage of assistive devices can help in mitigating these issues and enable the patients to communicate effectively. However, due to lack of large scale studies of linguistic deficits in aphasia, research on such assistive technology is relatively limited. In this work, we present two contributions that aim to re-initiate research and development in this field. Firstly, we propose a model that uses linguistic features from small scale studies on aphasia patients and generates large scale datasets of synthetic aphasic utterances from grammatically correct datasets. We show that the mean length of utterance, the noun/verb ratio, and the simple/complex sentence ratio of our synthetic datasets correspond to the reported features of aphasic speech. Further, we demonstrate how the synthetic datasets may be utilized to develop assistive devices for aphasia patients. The pre-trained T5 transformer is fine-tuned using the generated dataset to suggest 5 corrected sentences given an aphasic utterance as input. We evaluate the efficacy of the T5 model using the BLEU and cosine semantic similarity scores. Affirming results with BLEU score of 0.827/1.00 and semantic similarity of 0.904/1.00 were obtained. These results provide a strong foundation for the concept that a synthetic dataset based on small scale studies on aphasia can be used to develop effective assistive technology.

基于变压器的体系结构已在各种视觉域（最著名的图像和视频）中变得更具竞争力。虽然先前的工作已经孤立地研究了这些模式，但拥有一个共同的体系结构表明，人们可以训练单个统一模型以多种视觉方式。事先尝试进行统一建模通常使用针对视觉任务量身定制的体系结构，或与单个模态模型相比获得较差的性能。在这项工作中，我们表明可以使用蒙版的自动编码来在图像和视频上训练简单的视觉变压器，而无需任何标记的数据。该单个模型学习了与图像和视频基准上的单模式表示相当或更好的视觉表示，同时使用了更简单的体系结构。特别是，我们的单一预算模型可以进行审核，以在ImageNet上获得86.5％的速度，而在挑战性的事物V2视频基准测试中，可以实现75.3％的范围。此外，可以通过丢弃90％的图像和95％的视频补丁来学习该模型，从而实现非常快速的训练。

由于检测数据集的规模小，当前对象探测器的词汇量受到限制。另一方面，图像分类器的原因是大约更大的词汇表，因为他们的数据集更大，更容易收集。我们提出守则，只需在图像分类数据上培训检测器的分类器，从而扩展了探测器的词汇量到数万个概念。与现有工作不同，拒绝不会根据模型预测将图像标签分配给框，使其更容易实现和兼容一系列检测架构和骨架。我们的结果表明，即使没有箱子注释，否则差异也能产生出色的探测器。它优于开放词汇和长尾检测基准的事先工作。拒绝为所有类和8.3地图提供了2.4地图的增益，用于开放词汇LVIS基准测试中的新型类。在标准的LVIS基准测试中，守护者达到41.7地图所有课程和41.7地图以获得罕见课程。我们首次培训一个探测器，其中包含所有二十一千类的ImageNet数据集，并显示它在没有微调的情况下推广到新数据集。代码可在https://github.com/facebookresearch/dorm提供。

我们发现Mask2Former还可以在视频实例分段上实现最先进的性能，而无需修改架构，丢失甚至培训管道。在本报告中，我们通过直接预测3D分段卷来显示通用图像分割体系结构通过直接预测3D分段卷来概括到视频分段。具体而言，Mask2Former在Youtubevis-2021上为Youtubevis-2019和52.6 AP设置了新的60.4 AP最先进的。鉴于其在图像分割中的多功能性，我们认为蒙版2格相符也能够处理视频语义和Panoptic分割。我们希望这将使最先进的视频分段研究更可访问，并更加关注设计通用图像和视频分段架构。

图像分割是关于使用不同语义的分组像素，例如类别或实例成员身份，其中每个语义选择定义任务。虽然只有每个任务的语义不同，但目前的研究侧重于为每项任务设计专业架构。我们提出了蒙面关注掩模变压器（Mask2Former），这是一种能够寻址任何图像分段任务（Panoptic，实例或语义）的新架构。其关键部件包括屏蔽注意，通过限制预测掩模区域内的横向提取局部特征。除了将研究工作减少三次之外，它还优于四个流行的数据集中的最佳专业架构。最值得注意的是，Mask2Former为Panoptic semonation（Coco 57.8 PQ）设置了新的最先进的，实例分段（Coco上50.1 AP）和语义分割（ADE20K上的57.7 miou）。

The previous fine-grained datasets mainly focus on classification and are often captured in a controlled setup, with the camera focusing on the objects. We introduce the first Fine-Grained Vehicle Detection (FGVD) dataset in the wild, captured from a moving camera mounted on a car. It contains 5502 scene images with 210 unique fine-grained labels of multiple vehicle types organized in a three-level hierarchy. While previous classification datasets also include makes for different kinds of cars, the FGVD dataset introduces new class labels for categorizing two-wheelers, autorickshaws, and trucks. The FGVD dataset is challenging as it has vehicles in complex traffic scenarios with intra-class and inter-class variations in types, scale, pose, occlusion, and lighting conditions. The current object detectors like yolov5 and faster RCNN perform poorly on our dataset due to a lack of hierarchical modeling. Along with providing baseline results for existing object detectors on FGVD Dataset, we also present the results of a combination of an existing detector and the recent Hierarchical Residual Network (HRN) classifier for the FGVD task. Finally, we show that FGVD vehicle images are the most challenging to classify among the fine-grained datasets.

t-SNE remains one of the most popular embedding techniques for visualizing high-dimensional data. Most standard packages of t-SNE, such as scikit-learn, use the Barnes-Hut t-SNE (BH t-SNE) algorithm for large datasets. However, existing CPU implementations of this algorithm are inefficient. In this work, we accelerate the BH t-SNE on CPUs via cache optimizations, SIMD, parallelizing sequential steps, and improving parallelization of multithreaded steps. Our implementation (Acc-t-SNE) is up to 261x and 4x faster than scikit-learn and the state-of-the-art BH t-SNE implementation from daal4py, respectively, on a 32-core Intel(R) Icelake cloud instance.

Targeted syntactic evaluations of language models ask whether models show stable preferences for syntactically acceptable content over minimal-pair unacceptable inputs. Most targeted syntactic evaluation datasets ask models to make these judgements with just a single context-free sentence as input. This does not match language models' training regime, in which input sentences are always highly contextualized by the surrounding corpus. This mismatch raises an important question: how robust are models' syntactic judgements in different contexts? In this paper, we investigate the stability of language models' performance on targeted syntactic evaluations as we vary properties of the input context: the length of the context, the types of syntactic phenomena it contains, and whether or not there are violations of grammaticality. We find that model judgements are generally robust when placed in randomly sampled linguistic contexts. However, they are substantially unstable for contexts containing syntactic structures matching those in the critical test content. Among all tested models (GPT-2 and five variants of OPT), we significantly improve models' judgements by providing contexts with matching syntactic structures, and conversely significantly worsen them using unacceptable contexts with matching but violated syntactic structures. This effect is amplified by the length of the context, except for unrelated inputs. We show that these changes in model performance are not explainable by simple features matching the context and the test inputs, such as lexical overlap and dependency overlap. This sensitivity to highly specific syntactic features of the context can only be explained by the models' implicit in-context learning abilities.

Long-term OCR services aim to provide high-quality output to their users at competitive costs. It is essential to upgrade the models because of the complex data loaded by the users. The service providers encourage the users who provide data where the OCR model fails by rewarding them based on data complexity, readability, and available budget. Hitherto, the OCR works include preparing the models on standard datasets without considering the end-users. We propose a strategy of consistently upgrading an existing Handwritten Hindi OCR model three times on the dataset of 15 users. We fix the budget of 4 users for each iteration. For the first iteration, the model directly trains on the dataset from the first four users. For the rest iteration, all remaining users write a page each, which service providers later analyze to select the 4 (new) best users based on the quality of predictions on the human-readable words. Selected users write 23 more pages for upgrading the model. We upgrade the model with Curriculum Learning (CL) on the data available in the current iteration and compare the subset from previous iterations. The upgraded model is tested on a held-out set of one page each from all 23 users. We provide insights into our investigations on the effect of CL, user selection, and especially the data from unseen writing styles. Our work can be used for long-term OCR services in crowd-sourcing scenarios for the service providers and end users.