Add files using upload-large-folder tool

.gitattributes

@@ -16395,3 +16395,92 @@ parse/dev/XSEBx0iSjFQ/XSEBx0iSjFQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
 parse/dev/Ms6QZafNv01/Ms6QZafNv01_layout.pdf filter=lfs diff=lfs merge=lfs -text
 parse/dev/Ms6QZafNv01/Ms6QZafNv01_span.pdf filter=lfs diff=lfs merge=lfs -text
 parse/dev/3FvF1db-bKT/3FvF1db-bKT_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/YUEFlzlG_0c/YUEFlzlG_0c_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/YUEFlzlG_0c/YUEFlzlG_0c_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1eqOnNYDH/H1eqOnNYDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1eqOnNYDH/H1eqOnNYDH_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1eqOnNYDH/H1eqOnNYDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/rkEFLFqee/rkEFLFqee_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/rkEFLFqee/rkEFLFqee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/rkEFLFqee/rkEFLFqee_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SyxnvsAqFm/SyxnvsAqFm_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SyxnvsAqFm/SyxnvsAqFm_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SyxnvsAqFm/SyxnvsAqFm_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/ryloogSKDS/ryloogSKDS_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/ryloogSKDS/ryloogSKDS_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/ryloogSKDS/ryloogSKDS_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/i_Q1yrOegLY/i_Q1yrOegLY_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/i_Q1yrOegLY/i_Q1yrOegLY_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/i_Q1yrOegLY/i_Q1yrOegLY_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJxstlHFPH/SJxstlHFPH_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJxstlHFPH/SJxstlHFPH_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJxstlHFPH/SJxstlHFPH_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HJBhEMbRb/HJBhEMbRb_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HJBhEMbRb/HJBhEMbRb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HJBhEMbRb/HJBhEMbRb_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SkEqro0ctQ/SkEqro0ctQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SkEqro0ctQ/SkEqro0ctQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SkEqro0ctQ/SkEqro0ctQ_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/zv-typ1gPxA/zv-typ1gPxA_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/zv-typ1gPxA/zv-typ1gPxA_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/zv-typ1gPxA/zv-typ1gPxA_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJlhPMWAW/SJlhPMWAW_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJlhPMWAW/SJlhPMWAW_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJlhPMWAW/SJlhPMWAW_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/CoJibBRjPXQ/CoJibBRjPXQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/CoJibBRjPXQ/CoJibBRjPXQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/CoJibBRjPXQ/CoJibBRjPXQ_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1e_cC4twS/H1e_cC4twS_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1e_cC4twS/H1e_cC4twS_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1e_cC4twS/H1e_cC4twS_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJl_VnR9Km/BJl_VnR9Km_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJl_VnR9Km/BJl_VnR9Km_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJl_VnR9Km/BJl_VnR9Km_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/Skgb5h4KPH/Skgb5h4KPH_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/Skgb5h4KPH/Skgb5h4KPH_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/Skgb5h4KPH/Skgb5h4KPH_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/iklDP-v1sgi/iklDP-v1sgi_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/iklDP-v1sgi/iklDP-v1sgi_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/iklDP-v1sgi/iklDP-v1sgi_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/fIn4wLS2XzU/fIn4wLS2XzU_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/fIn4wLS2XzU/fIn4wLS2XzU_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/fIn4wLS2XzU/fIn4wLS2XzU_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1xk8jAqKQ/H1xk8jAqKQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1xk8jAqKQ/H1xk8jAqKQ_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/H1xk8jAqKQ/H1xk8jAqKQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/yvQKLaqNE6M/yvQKLaqNE6M_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/yvQKLaqNE6M/yvQKLaqNE6M_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/yvQKLaqNE6M/yvQKLaqNE6M_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJE-4xW0W/BJE-4xW0W_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJE-4xW0W/BJE-4xW0W_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJE-4xW0W/BJE-4xW0W_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/Hkg5lAEtvS/Hkg5lAEtvS_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/Hkg5lAEtvS/Hkg5lAEtvS_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/Hkg5lAEtvS/Hkg5lAEtvS_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/OU98jZWS3x_/OU98jZWS3x__origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/OU98jZWS3x_/OU98jZWS3x__layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/OU98jZWS3x_/OU98jZWS3x__span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJbD_Pqlg/BJbD_Pqlg_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJbD_Pqlg/BJbD_Pqlg_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJbD_Pqlg/BJbD_Pqlg_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJeeL04KvH/SJeeL04KvH_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJeeL04KvH/SJeeL04KvH_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/SJeeL04KvH/SJeeL04KvH_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HJx8HANFDH/HJx8HANFDH_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HJx8HANFDH/HJx8HANFDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HJx8HANFDH/HJx8HANFDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HkpYwMZRb/HkpYwMZRb_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HkpYwMZRb/HkpYwMZRb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/HkpYwMZRb/HkpYwMZRb_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BkgnhTEtDS/BkgnhTEtDS_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BkgnhTEtDS/BkgnhTEtDS_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BkgnhTEtDS/BkgnhTEtDS_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/ByG8A7cee/ByG8A7cee_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/ByG8A7cee/ByG8A7cee_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/ByG8A7cee/ByG8A7cee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJj6qGbRW/BJj6qGbRW_span.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJj6qGbRW/BJj6qGbRW_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/BJj6qGbRW/BJj6qGbRW_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/rkeZIJBYvr/rkeZIJBYvr_layout.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/rkeZIJBYvr/rkeZIJBYvr_origin.pdf filter=lfs diff=lfs merge=lfs -text
+parse/train/rkeZIJBYvr/rkeZIJBYvr_span.pdf filter=lfs diff=lfs merge=lfs -text

parse/dev/_LceCyuVcH/_LceCyuVcH.md

@@ -0,0 +1,269 @@
+# Language Models with Image Descriptors are Strong Few-Shot Video-Language Learners
+
+Zhenhailong Wang1∗, Manling $\mathbf { L i } ^ { 1 }$ ∗, Ruochen $\mathbf { X } \mathbf { u } ^ { 2 }$ , Luowei Zhou2†, Jie Lei3, Xudong Lin4, Shuohang $\mathbf { W a n g } ^ { 2 }$ , Ziyi $\mathbf { Y a n g } ^ { 2 }$ , Chenguang $\mathbf { Z } \mathbf { h } \mathbf { u } ^ { 2 }$ , Derek Hoiem1, Shih-Fu Chang4, Mohit Bansal3, Heng Ji1 1UIUC 2MSR 3UNC 4Columbia University {wangz3,hengji}@illinois.edu
+
+# Abstract
+
+The goal of this work is to build flexible video-language models that can generalize to various video-to-text tasks from few examples. Existing few-shot video-language learners focus exclusively on the encoder, resulting in the absence of a video-totext decoder to handle generative tasks. Video captioners have been pretrained on large-scale video-language datasets, but they rely heavily on finetuning and lack the ability to generate text for unseen tasks in a few-shot setting. We propose VidIL, a few-shot Video-language Learner via Image and Language models, which demonstrates strong performance on few-shot video-to-text tasks without the necessity of pretraining or finetuning on any video datasets. We use image-language models to translate the video content into frame captions, object, attribute, and event phrases, and compose them into a temporal-aware template. We then instruct a language model, with a prompt containing a few in-context examples, to generate a target output from the composed content. The flexibility of prompting allows the model to capture any form of text input, such as automatic speech recognition (ASR) transcripts. Our experiments demonstrate the power of language models in understanding videos on a wide variety of video-language tasks, including video captioning, video question answering, video caption retrieval, and video future event prediction. Especially, on video future event prediction, our few-shot model significantly outperforms state-of-the-art supervised models trained on large-scale video datasets. Code and processed data are publicly available for research purposes at https://github.com/MikeWangWZHL/VidIL.
+
+# 1 Introduction
+
+One major gap between artificial intelligence and human intelligence lies in their abilities to generalize and perform well on new tasks with limited annotations. Recent advances in large-scale pre-trained generative language models [45, 6, 71, 24] have shown promising few-shot capabilities [72, 43, 63] in understanding natural language. However, few-shot video-language understanding is still in its infancy. A particular limitation of most recent video-language frameworks [28, 21, 61, 68, 67, 25, 64, 34] is that they are encoder-only, which means they do not have the ability to generate text from videos for purposes such as captioning [62, 57], question answering [60], and future prediction [23]. Meanwhile, unified video-language models [36, 49] that are capable of language decoding still rely heavily on finetuning using a large number of manually annotated video-text pairs, therefore cannot adapt quickly to unseen tasks. Few-shot video-to-text decoding is challenging because the natural language supervision for learning video-language representation is typically based on subtitles and automatic speech recognition (ASR) transcripts [39, 68], which differ significantly from downstream tasks in terms of distribution and may have poor semantic alignment across vision and text modalities.
+
+We propose to address this problem by harnessing the few-shot power of frozen large-scale language models, such as InstructGPT [40]. Our inspiration is derived from the fact that humans are excellent visual storytellers [15], with the ability to piece together a coherent story from a few isolated images. To mimic this, we propose VidIL, a few-shot Video-language Learner via Image and Language models, to use image models to provide information about the visual content in the video (as well as optionally use ASR to represent speech), and then we instruct language models to generate a video-based summary, answer, or other target output for diverse video-language tasks.
+
+The main challenge of understanding videos is that, videos contain rich semantics and temporal content at multiple granularities. Unlike static images which depict objects, attributes and events in a snapshot, the temporal dimension of videos further conveys the state changes of the objects, actions, and events. For example, in Figure 1, the individual frame captions of the video clip only describe static visual features such as "a person holding a green object in hand". In contrast, a correct video-level description would be "a woman makes realistic looking leaves and flowers for a cake", which involves reasoning over a collection of objects and events that occur at different timestamps in the video clip, such as "cake decorating" and "flowered design". Hence, to inform video-level description and queries, we need to represent all of this information and its temporal ordering.
+
+![](images/b5f930daf9032229a0d0a4e399381948a20cacaccc02903750ded88494988f0e.jpg)  
+Figure 1: Multiple levels of information in videos.
+
+To address the unique challenges of videos, we propose to decompose a video into three levels: the video output, frame captions, and visual tokens (including objects, events, attributes). One major benefit from this hierarchical video representation is that we can separate the visual and temporal dimensions of a video. We leverage frozen image-language foundational models at lower levels to collect salient visual features from the sparsely sampled frames. Specifically, we first leverage a pretrained image-language contrastive model CLIP [44] to perform visual tokenization, based on the similarity score between frames and tokens of objects, events and attributes. The tokenization is done under the guidance of semantics role labeling [14], which provides us with candidate events with involved objects and related attributes. Next, in order to capture the overall semantics at the frame level, we employ the pretrained image captioner in the image-language model BLIP [26] to obtain frame captions. Finally, we instruct a pretrained large language model using in-context learning [40, 13, 51, 48] to interpret visual tokens and frame captions into the target textual output. In detail, we temporally order visual tokens and frame captions using specially designed prompts such as “First...Then...Finally”, to instruct the pretrained language model to track the changes of objects, events, attributes and frame semantics along the temporal dimension.
+
+Without pretraining or finetuning on any video datasets, we show that our approach outperforms both video-language and image-language state-of-the-art baselines on few-shot video captioning and question answering tasks. Moreover, on video-language event prediction, our approach significantly outperforms fully-supervised models while using only 10 labeled examples. We further demonstrate that our generative model can benefit broader video-language understanding tasks, such as text-video retrieval, via pseudo label generation. Additionally, we show that our model is highly flexible in adding new modalities, such as ASR transcripts.
+
+# 2 Related Work
+
+# 2.1 Image-Language Models and Their Applications on Video-Language Tasks
+
+Large-scale image-language pretraining models optimize image-text matching through contrastive learning [44, 17] and multimodal fusion [65, 27, 58, 66, 35, 52, 8, 29, 73, 70, 18, 16]. Recently,
+
+BLIP [26] proposes a bootstrapping image-language pretraining framework with a captioner and a filterer which has shown promising performance on various image-language tasks. However, video-language pretraining [25, 36, 28, 38, 3, 1, 42, 33] is still hindered by noisy and domain-specific video datasets [74, 22, 39]. Naturally, researchers start to explore transferring the rich knowledge from image models to videos. Different from the traditional way of representing videos by 3D dense features [12], recent work [21, 25] proves that sparse sampling is an effective way to represent videos, which facilitates applying pre-trained image-language models to video-language tasks [37, 11]. Specifically, the image-language model BLIP [26] sets new state-of-the-art on zero-shot retrieval-style video-language tasks, such as video retrieval and video question answering. However, for generationstyle tasks such as domain-specific video captioning, video-language model UniVL [36] still leads the performance but highly rely on fine-tuning. In this work, we extend the idea of leveraging image-language models to a wide variety of video-to-text generation tasks. We further connect imagelanguage models with language models which empowers our model with strong generalization ability. We show that the knowledge from both image-language pretraining and language-only pretraining can benefit video-language understanding in various aspects.
+
+# 2.2 Unifying MultiModal Tasks with Language Models
+
+The community has paid much attention to connecting different modalities with a unified representation recently. Text-only generation models, such as T5 [46], have been extended to vision-language tasks by text generation conditioned on visual features [9, 53, 50, 75, 55]. In order to fully leverage the generalization power from pretained language models, [63] represents images using text in a fully symbolic way. [32] includes more modalities such as video and audio, but requires annotated video-text data to jointly training the language model with the video and audio tokenizer. In this work, we propose a temporal-aware hierarchical representation for describing a video textually. To our knowledge, we are the first work to leverage prompting a frozen language model for tackling few-shot video-language tasks with a unified textual representation. Concurrent work Socratic [69] uses a zero-shot language-based world-state history to represent long videos with given time stamps, while our model can quickly adapt to different video and text distributions with few examples. Furthermore, we show that by injecting temporal markers to the prompt we can make a pre-trained language model understand fine-grained temporal dynamics in video events. Compared with the concurrent work Flamingo [2], which requires dedicated vision-language post-pretraining, our framework does not require to pretrain or finetune on any video data. Our framework is simple and highly modulated where all the components are publicly available. Additionally, our framework is more flexible on adding new modalities, e.g., automatic speech recognition, without the need for complex redesigning.
+
+# 3 Method
+
+We propose a hierarchical video representation framework which decomposes a video into three levels, i.e., visual token level, frame level and video level. The motivation is to separate the spatial and temporal dimension of a video in order to leverage image-language and language-only foundation models, such as CLIP [44] and GPT-3 [6]. All three levels use a unified textual representation which enables us to leverage the powerful few-shot ability from pretrained language models.
+
+# 3.1 Frame Level: Image Captioning
+
+Following [21] we first perform sparse sampling to obtain several video frames. Unless otherwise specified, we sample 4 frames for frame level and 8 frames for visual token level. We then feed each frame into a pre-trained image-language model to obtain frame level captions. An example can be found in the blue part of Figure 2. In our experiments, we use BLIP [26], a recent image-language framework containing both image-grounded encoder and decoder, for generating frame captions. We follow [26] to do both captioning and filtering on each frame. However, as mentioned in Section 1, videos contain rich semantics and temporal contents at multiple granularities. It is not enough to generate video-level target text such as video captions solely based on frame captions. Thus, we further perform visual tokenization for each frame to capture features at a finer granularity.
+
+![](images/c144de0cb595e241110b349a7da85b61029ef363498acc4c08b0c7647b08efae.jpg)  
+Figure 2: Overview of VidIL framework. We represent a video in a unified textural representation containing three semantic levels: visual token level, frame level, and video level. At visual token level, we extract salient objects, events, attributes for each sampled frame. At frame level, we perform image captioning and filtering. At video level, we construct video representation by aggregating the visual tokens, frame captions and other text modalities such as ASR, using a few-shot temporalaware prompt. We then feed the prompt to a pre-trained language model together with task-specific instructions to generate target text for a variety of video-language tasks. Examples of the full prompt for different tasks can be found in Appendix ??.
+
+# 3.2 Visual Token Level: Structure-Aware Visual Tokenization
+
+At this level, we aim to extract the textual representations of salient visual token types, such as objects, events and attributes. We found that pre-defined classes for classification, such as those in ImageNet [10], are far from enough for covering the rich semantics in open-domain videos. Thus, instead of using classification-based methods for visual tokenization as in previous work [32, 63], we adopt a retrieval-based visual tokenization approach by leveraging pre-trained contrastive imagelanguage models. Given a visual token vocabulary which contains all candidate object, event, and attribute text phrases, we compute the image embedding of a frame and the text embeddings of the candidate visual tokens using a contrastive multi-modal encoder, CLIP [44]. We then select top 5 visual tokens per frame based on the cosine similarity of the image and text embeddings. An example of the extracted object tokens can be found in the green part of Figure 2.
+
+Unlike in images where objects and attributes already cover most visual features, events are more informative in videos. In order to discover events from video frames, we construct our own event vocabulary by extracting event structures from Visual Genome [19] synsets3 using Semantic Role Labeling. Specifically, we first select the phrases that contain at least one verb and one argument as events. Then we remove highly similar events based on their sentence similarity using SentenceBERT [47] embeddings. For object vocabulary, we adopt OpenImage [20] full classes $( { \sim } 2 0 \mathbf { k } )$ , instead of using the visually groundable subset $( \sim 6 0 0 )$ as in concurrent work [69]. We found that using large but noisy vocabulary is more effective than using small but clean vocabulary in our retrieval-based setting with CLIP. For attribute vocabulary, we adopt visual genome attribute synset. In Section 4.6, we provide ablation study on the impact of different types of visual tokens. The statistics of visual token vocabulary can be found in Appendix Table ??.
+
+![](images/331592c36e0944e6b5f6973236f3089e33d229d668d68f0224a0f67fdb784fba.jpg)  
+Figure 3: Temporal-aware prompt successfully distinguishes the Sunset and Sunrise scenarios based on the temporal ordering change of objects and frame captions, while the static prompt fails.
+
+# 3.3 Video Level: Temporal-Aware Few-shot Prompting
+
+Once we obtain the textual representation from frame level and visual token level, the final step is to put the pieces together to generate a video level target text. The goal is to build a model that can be quickly adapted to any video-to-text generation task with only a few examples. To this end, we propose to leverage large-scale pre-trained language models, such as GPT-3 [6], with a temporal-aware few-shot prompt. As shown in Figure 2, our framework can be readily applied to various video-to-text generation tasks, such as video captioning and video question answering, with a shared prompt template. The proposed prompting strategy enables a language model to attend to the lower level visual information as well as taking into account the temporal ordering.
+
+Here, we use the video captioning task depicted in Figure 2 to illustrate the details. The few-shot prompt consists of three parts: instruction, few-shot context, and task query. The instruction is a concise description of the generation task, e.g., "Generate a video caption based on the objects, events, attributes and frame captions. Example:", which is proved to be effective in zero-shot and few-shot settings [6, 59]. The few-shot context contains the selected in-context examples as well as the test video instance. Each video instance is represented by the aggregated visual tokens4, e.g., "Objects: First, bath toy. Then,..." the frame captions, such as "Frame Captions: First, a toddler playing in a bathtub filled with toys. Then,...", and the ASR inputs if available, e.g., "Subtitle:<ASR Transcript>". Finally, the task query is a task-specific suffix indicating the target text format, e.g. "Video Caption:". For in-context examples (omitted here for simplicity), the task query is followed by ground truth annotation, while for the test instance, the generation starts at the end of the task query.
+
+Formally, we denote the instruction line as $\mathbf { t }$ , few-shot context as c, the task query as q, and the target text as $\mathbf { y }$ , where $\mathbf { y } = ( y _ { 1 } , y _ { 2 } , . . . , y _ { L } )$ . The generation of the next target token $y _ { l }$ can be modeled as:
+
+$$
+y _ { l } = \mathop { \arg \operatorname* { m a x } } _ { y } p ( y | \mathbf { s } , \mathbf { c } , \mathbf { q } , y _ { < l } )
+$$
+
+In order to capture the temporal dynamics between frames and visual tokens, we further propose to inject temporal markers to the prompt. As shown in the few-shot context in Figure 2, each visual token and frame caption is prefixed with a natural language phrase indicating its temporal ordering, e.g., "First,","Then,", and "Finally,". We found adding the temporal marker can make the language model conditioned on not only literal but also temporal information of the context. We show an example in Figure 3, where we compare our temporal-aware prompt with a static prompt on video captioning using InstructGPT. Again, the in-context examples are omitted here, which can be found in Appendix ??. In this example, the only difference between these two contexts is the ordering of the visual tokens and the frame captions. For the context on the left, where "sun moving" appears before "night sky", we are expected to see a story talking about sunset, while for the context on the right, we are expected to see sunrise. We can see the static prompt generates captions about sunset for both contexts, while the temporal-aware prompt can capture temporal ordering correctly and generate sunrise for the context on the right.
+
+# 4 Experiments
+
+# 4.1 Experimental Setup
+
+To comprehensively evaluate our model, we show results on four video-language understanding tasks in few-shot settings: video captioning, video question answering (QA), video-language event prediction, and text-video retrieval. We compare our approach with state-of-the-art approaches on five benchmarks, i.e, MSR-VTT [62], MSVD [7], VaTeX [57], YouCook2 [74], and VLEP [23]. The statistics of the datasets can be found in Table 1. For more details please refer to Appendix ??.
+
+Implementation Details. We use CLIP-L/146 as our default encoder for visual tokenization. We adopt BLIP captioning checkpoint7 finetuned on COCO [31] for frame captioning. We use InstructGPT [40] as our default language model for generating text conditioned on the few-shot prompt. To construct event vocabulary, we use the semantic role labeling model from AllenNLP8. The experiments are conducted on 2 NVIDIA V100 (16GB) GPUs. All few-shot
+
+Table 1: Statistics of datasets in our experiments   
+
+<table><tr><td>Dataset</td><td>Task</td><td>Split Count # train /#eval</td></tr><tr><td>MSR-VTT[62]</td><td>Captioning; QA</td><td>6,513 /2.990</td></tr><tr><td>MSR-VTT[62]</td><td>Retrieval</td><td>7,010 / 1,000</td></tr><tr><td>MSVD [7]</td><td>Question Answering</td><td>30,933 /13,157</td></tr><tr><td>VaTeX v1.15 [57]</td><td>Captioning; Retrieval</td><td>25,991/6.000</td></tr><tr><td>YouCook2[74]</td><td>Captioning</td><td>10,337 /3,492</td></tr><tr><td>VLEP [23]</td><td>Event Prediction</td><td>20,142/4,192</td></tr></table>
+
+finetuning on baselines and semi-supervised training are performed on 2 Nvidia V100 16G GPUs.
+
+In-context Example Selection. From our preliminary experiments, we find that the generation performance is sensitive to the quality of in-context examples. For example, for QA tasks such as MSVD-QA where the annotations are automatically generated, the <question, answer> pair in randomly selected in-context examples can be only weakly-correlated with the video context. Thus, instead of using a fixed prompt for each query, we dynamically filter out the irrelevant in-context examples. Specifically, given a randomly sampled $M .$ -shot support set from the training set, we select a subset of $N _ { ☉ }$ -shots as in-context examples based on their SentenceBERT [47] similarities with text queries. Furthermore, we reorder the selected examples in ascending order based on the similarity score to account for the recency bias [72] in large language models. For QA tasks, we choose the most relevant in-context examples by comparing with questions. While for captioning task, we compare with frame captions. If not otherwise specified, we use $M { = } I O$ and $N { = } 5$ , which we consider as 10-shot training.
+
+# 4.2 Few-shot Video Captioning
+
+We report BLEU-4 [41], ROUGE-L [30], METEOR [5], and CIDEr [54] scores on three video captioning benchmarks covering both open-domain (MSR-VTT, VaTeX) and domain-specific (YouCook2) videos. We compare with both state-of-the-art video captioner (UniVL [36]) and image captioner (BLIP [26]). In order to implement the BLIP baseline for few-shot video captioning, we extend the approach used for text-video retrieval evaluation in [26] to video-language training. Specifically, we concatenate the visual features of sampled frames and then feed them into the image-grounded text-encoder to compute the language modeling loss. This is equivalent to stitching the sampled frames into a large image and then feeding it to BLIP for image captioning. We found that this simple approach results in very strong baselines.
+
+As shown in Table 2, existing methods have strong bias on certain datasets. For example, UniVL performs well on YouCook2 but fails on MSR-VTT and VaTeX, while BLIP performs the opposite. This is because UniVL is pretrained on HowTo100M which favors instructional videos, i.e., YouCook2, while BLIP is pre-training on image-caption pairs which favors description-style captions, i.e., MSR-VTT and VaTeX. On the contrary, our model performs competitively on both open-domain and instructional videos, and significantly outperforms the baselines on the average CIDEr score across all three benchmarks. This indicates that by leveraging language models, we can maintain strong few-shot ability regardless of the video domain or the target caption distribution.
+
+Table 2: 10-shot video captioning results. ♠ indicates concurrent work. The reported Flamingo [2] results are using 16 shots. #VideoPT represents the number of videos used for pre-training. B-4, R-L, $M$ , $C$ represents BLEU-4, ROUGE-L, METEOR and CIDEr. Avg $C$ represents the average CIDEr score across all available benchmarks. ASR indicates whether the model has access to the ASR subtitles. $B L I P$ and $B L I P _ { c a p }$ use the pretrained checkpoint and the finetuned checkpoint on COCO captioning. All results are averaged over three random seeds.   
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">#VideOpr ASR</td><td rowspan="2"></td><td rowspan="2">B-4R-LM</td><td colspan="3">MSR-VTTCaption</td><td colspan="3">YouCook2 Caption C</td><td rowspan="2"></td><td colspan="3">VaTex Caption</td><td rowspan="2">Avg C</td></tr><tr><td></td><td></td><td>C</td><td>B-4R-LM</td><td></td><td></td><td>B-4 R-L M</td><td></td><td>C</td></tr><tr><td>Few-shot</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>UniVL</td><td>1.2M</td><td>No</td><td>2.1 22.5 9.5</td><td></td><td></td><td>3.6</td><td>3.3</td><td>25.3 11.6</td><td>34.1</td><td>1.7</td><td>15.7</td><td>8.0</td><td>2.1</td><td>13.3</td></tr><tr><td>BLIP</td><td>0</td><td>No</td><td>27.7 43.0 23.0 39.5</td><td></td><td></td><td></td><td>0.7</td><td>9.0 3.4</td><td>11.5</td><td>13.5</td><td>39.5</td><td>15.4</td><td>20.7</td><td>23.9</td></tr><tr><td>BLIPcap</td><td>0</td><td>No</td><td>21.648.022.7 30.2</td><td></td><td></td><td></td><td>3.7</td><td>8.6 3.8</td><td>9.4</td><td>20.7</td><td>41.5</td><td>17.4</td><td>28.9</td><td>22.8</td></tr><tr><td>VidIL(ours)</td><td>0</td><td>No</td><td>26.0 51.7 24.7 36.3</td><td></td><td></td><td></td><td>2.6</td><td>22.9 9.5</td><td>27.0</td><td>22.2</td><td></td><td>43.620.0</td><td>36.7</td><td>33.3</td></tr><tr><td>UniVL</td><td>1.2M</td><td>Yes</td><td>-</td><td></td><td>=</td><td>-</td><td>4.3</td><td>26.4 12.2</td><td>48.6</td><td>2.7</td><td>17.7</td><td>10.2</td><td>3.4</td><td>26.0</td></tr><tr><td>VidIL(ours)</td><td>0</td><td>Yes</td><td></td><td></td><td></td><td></td><td>10.7 35.9</td><td></td><td>19.4 111.6</td><td></td><td>23.2 44.2 20.6 38.9</td><td></td><td></td><td>75.3</td></tr><tr><td>Flamingo-3B(16)</td><td>27M</td><td>No</td><td></td><td></td><td></td><td></td><td></td><td></td><td>73.2</td><td></td><td></td><td></td><td>57.1</td><td></td></tr><tr><td>Flamingo-80B(16) 27M</td><td></td><td>No</td><td></td><td></td><td></td><td></td><td></td><td></td><td>84.2</td><td></td><td></td><td></td><td>62.8</td><td>=</td></tr><tr><td>Fine-tuning</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>UniVL</td><td>1.2M</td><td>No</td><td>42.0 61.0 29.0 50.1</td><td></td><td></td><td></td><td></td><td></td><td></td><td>[11.2 40.1 17.6 127.0|22.8 38.6 22.3 33.4</td><td></td><td></td><td></td><td>70.2</td></tr><tr><td>UniVL</td><td>1.2M</td><td>Yes</td><td>=</td><td></td><td></td><td></td><td></td><td></td><td></td><td>16.6 45.7 21.6176.823.7 39.3 22.7 35.6</td><td></td><td></td><td></td><td>106.2</td></tr></table>
+
+As discussed in Section 1, video captions describe the content in various semantic levels. The N-gram based metric may not fairly reflect the models’ performance in capturing the video-caption alignment. We further verify this hypothesis in Section 4.5. Thus, in addition to automatic metrics, we include qualitative examples illustrated in Figure 4. More examples are in Appendix ??.
+
+Additionally, for most existing methods and also concurrent work, e.g., Flamingo [2], adding a new modality often requires a dedicated model redesign or retraining. However, the nature of our framework, where we use a unified textual representation for each level, makes it highly flexible for incorporating new modalities. As shown in row 6 in Table, our model can effectively utilize extra information from ASR to obtain significantly better few-shot performance on certain datasets such as YouCook2.
+
+![](images/42488e5e146be037d7c57abcc48a78f8af1cf75e6517bc50d3e1b32bfd715b6a.jpg)
+
+Figure 4: Qualitative examples on video captioning. Grey boxes contain part of the video representation from our model. Blue boxes contain caption generation from different models. Green boxes contain ground truth annotations. Bold green text highlights the correct information that is not captured in baseline outputs which can be reasoned from our visual tokens and frame captions.
+
+# 4.3 Few-shot Video Question Answering
+
+We compare the test accuracy of our approach with few-shot pretrained BLIP, $\mathsf { B L I P } _ { V Q A }$ [26], and concurrent work Flamingo [2] on two video question answering benchmarks, MSR-VTT_QA and MSVD_QA. $\mathsf { B L I P } _ { V Q A }$ represents finetuned BLIP on VQA [4] dataset, which is the previous SOTA on zero/few-shot video question answering. In order to have fairer comparison with $\mathsf { B L I P } _ { V Q A }$ , we reduce the shot number to 5 and report the average accuracy on three sets of randomly selected 5-shot examples. As shown in Table 3, our method outperforms previous SOTA by a large margin. Comparing with concurrent work Flamingo, which is post-pretrained on a large number of video-text data, our model is training-free and did not observe any video data. However, with only imagelanguage and language-only knowledge, our 5-shot model is able to outperform 8-shot Flamingo-3B and achieve on-par performance with 4-shot Flamingo-80B.
+
+Table 3: Video QA results. $\mathsf { B L I P } _ { V Q A }$ is finetuned on VQA [4]. ♠ indicates concurrent work. PT, FT indicates pretraining and finetuning.   
+
+<table><tr><td>Method</td><td>#videoPT</td><td>#videoFT</td><td>MSR-VTT</td><td>MSVD</td></tr><tr><td>BLIP</td><td>0</td><td>O-shot</td><td>0.55</td><td>0.45</td></tr><tr><td>BLIP</td><td>0</td><td>5-shot</td><td>0.84</td><td>0.53</td></tr><tr><td>BLIPvQA [26]</td><td>0</td><td>O-shot</td><td>19.2</td><td>35.2</td></tr><tr><td>VidIL(ours)</td><td>0</td><td>5-shot</td><td>21.2</td><td>39.1</td></tr><tr><td>Flamingo-3B [2]</td><td>27M</td><td>4-shot</td><td>14.9</td><td>33.0</td></tr><tr><td>Flamingo-3B_[2]</td><td>27M</td><td>8-shot</td><td>19.6</td><td>37.0</td></tr><tr><td>Flamingo-80B [2]</td><td>27M</td><td>4-shot</td><td>23.9</td><td>41.7</td></tr><tr><td>Flamingo-80B [2]</td><td>27M</td><td>8-shot</td><td>27.6</td><td>45.5</td></tr><tr><td>ALPRO [25]</td><td>2M</td><td>full-shot</td><td>42.1</td><td>45.9</td></tr></table>
+
+Table 4: Accuracy $( \% )$ on VLEP hidden test set.   
+
+<table><tr><td>Method</td><td>#videoFT</td><td>Acc</td></tr><tr><td>VLEP [23] MERLOT[68]</td><td>20142 20142</td><td>67.5 68.4</td></tr><tr><td>VidIL(ours)</td><td>10-shot</td><td>72.0</td></tr><tr><td>Human</td><td>-</td><td>90.5</td></tr></table>
+
+# 4.4 Few-shot Video-Language Event Prediction
+
+In this section, we show that our model not only can answer questions about the video visual features but also answering "What is more likely to happen next?". Given a video with associated subtitle transcript as premise, the video-language event prediction (VLEP) task is to predict the most likely future event. The original VLEP [23] paper formulates the problem as a binary classification problem where the model will be chosen from two possible future event candidates. Instead, we formulate this problem as another video-to-text generation problem to fit into our framework. Figure 5 depicts an example with the same format as in Figure 2. Similar to the evaluation setting in QA, the generated free-form text will first be mapped to one of the two candidate answers using SentenceBert [47], and then calculate the accuracy. In Table 4, we report accuracy on the hidden test set of VLEP [23]. To our surprise, our 10-shot model outperforms state-of-the-art fully-supervised baseline, i.e., MERLOT [68], by a large margin $( \sim 4 \% )$ . This shows that our model has strong few-shot ability not only on videolanguage understanding but also on prediction. Since event prediction tasks rely heavily on temporal ordering, we show that with the proposed temporal-aware prompting, language models can be guided to capture temporal dynamics between historical and future events.
+
+![](images/337717ccf8591fcb87cdc2bcb90b9982ea93bafdcb263be40f57ceb4066b57ea.jpg)  
+Figure 5: Prompt for VLEP task.
+
+# 4.5 Semi-supervised Text-Video Retrieval
+
+In addition to video-to-text generation tasks, we show that a broader range of video-language tasks can benefit from our few-shot video captioner from a data perspective. Here, we consider a lowbudget semi-supervised setting where we only have a few labeled video-caption pairs and a large amount of unlabeled videos. The idea is to leverage our video captioner to generate pseudo labels for training any given vision-language models. As a case study, we evaluate on two text-video retrieval benchmarks, i.e., MSR-VTT and VaTeX. We use greedy decoding to generate pseudo caption for each video in the training set. We then train an identical base model, i.e., BLIP, using different pseudo labeled data as well as ground truth annotations. We report Recall $@$ 1 and 5 for both video-to-text and text-to-video retrieval. Table 5 shows that through training on our pseudo labels, we can achieve significant improvements compared with zero-shot BLIP. We also show that the performance gain is not simply a result of training on more data, since finetuning on the pseudo labels generated by other baselines (UniVL, BLIP) is less effective and can even hurt the performance. Furthermore, on MSR-VTT Recall $@$ 5 we can even achieve comparable performance against BLIP model finetuned on full ground truth annotations.
+
+Table 5: Semi-supervised text-video retrieval with 10 labeled examples. $\mathrm { V _ { l a b e l } }$ or $\mathrm { \Delta V _ { u n l a b e l } }$ are the number of labeled and unlabeled videos, respectively. $t \_ R I$ and $t \_ R$ denote video-to-text Recall $@ 1$ and 5. $\nu \_ R I$ and $\nu \_ R 5$ denote text-to-video Recall $@ 1$ and 5.   
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Pseudo Label</td><td colspan="5">MSR-VTTRetrieval</td><td colspan="5">VaTexRetrieval</td></tr><tr><td>Vlabel/Vunlabel t_R1t_R5v_R1</td><td></td><td></td><td></td><td>v_R5</td><td>Vlabel/Vunlabel t_R1t_R5v_R1</td><td></td><td></td><td></td><td>v_R5</td></tr><tr><td>BLIP</td><td></td><td></td><td>33.2</td><td>57.2</td><td>40.5</td><td>62.8</td><td></td><td>28.2</td><td>53.4</td><td>34.0</td><td>58.6</td></tr><tr><td>BLIP</td><td>UniVL</td><td>10 /7010</td><td>33.1</td><td>57.3</td><td>33.6</td><td>57.7</td><td>10 /22685</td><td>25.5</td><td>47.7</td><td>26.1</td><td>49.1</td></tr><tr><td>BLIP</td><td>BLIP</td><td>10/7010</td><td>35.6</td><td>60.8</td><td>39.8</td><td>60.4</td><td>10/22685</td><td>26.3</td><td>50.5</td><td>29.3</td><td>53.6</td></tr><tr><td>BLIP</td><td>BLIPcap</td><td>10/7010</td><td>35.3</td><td>58.0</td><td>39.1</td><td>63.3</td><td>10/22685</td><td>23.9</td><td>46.8</td><td>27.5</td><td>49.7</td></tr><tr><td>BLIP</td><td>VidIL(ours)</td><td>10/7010</td><td>39.6</td><td>64.5</td><td>40.8</td><td>65.2</td><td>10 /22685</td><td>33.3</td><td>59.1</td><td>33.7</td><td>59.5</td></tr><tr><td>BLIP</td><td>Ground Truth</td><td>7010/0</td><td>43.6</td><td>66.2</td><td>43.1</td><td>67.2</td><td>22685/0</td><td>40.1</td><td>66.4</td><td>40.1</td><td>66.6</td></tr><tr><td>ALPRO [25]</td><td>Ground Truth</td><td>140200/0</td><td>32.0</td><td>60.6</td><td>33.9</td><td>60.7</td><td></td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DRL [56]</td><td>Ground Truth</td><td>180000/0</td><td>54.1</td><td>77.4</td><td>52.9</td><td>78.5</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr></table>
+
+Another interesting observation is that, compared with the video captioning results in Table 2, we found that the gain of our model over baselines on text-video retrieval is more visible than on captioning. A key factor in performing well on text-video retrieval tasks is to learn a good video-text multi-modal alignment. This result shows that our pseudo labels capture richer video-text alignment that can benefit the retrieval-style downstream task. The N-gram based generation metrics, e.g., BLEU, may not be able to fully reflect the alignment information, due to the variety of semantic levels in video captions. Furthermore, from a data perspective, our video captioner can be viewed as a data augmentation tool which is capable of generating or augmenting any open-domain videolanguage pretraining datasets with minimal human effort. As a result, we can potentially improve video-language pretraining by constructing a cleaner and more diverse video-text corpus.
+
+Table 6: Impact of visual tokens and temporal dimension.   
+
+<table><tr><td colspan="2">Video Representation</td><td>Avg↑ Std↓</td><td></td></tr><tr><td rowspan="3">Visual Token</td><td>Frame Frame+Object</td><td>39.6 40.3</td><td>3.7 2.9</td></tr><tr><td>Frame+Object+Event</td><td>39.9</td><td>2.8</td></tr><tr><td>Frame+Object+Attibute Frame+Object+Event+Attribute</td><td>40.9 40.8</td><td>2.9 2.4</td></tr><tr><td>Temporal</td><td></td><td>Reduce to one frame Reverse temporal order</td><td>38.5 40.7</td><td>2.4 1.7</td></tr></table>
+
+Table 7: Impact of shot selection. #ICE indicates the number of in-context examples in the prompt. Details of in-context example selection are in the Appendix.   
+
+<table><tr><td>#shot</td><td colspan="3">w/o selection</td><td colspan="3">w/ selection</td></tr><tr><td></td><td>#ICE</td><td>Avg↑</td><td>Std</td><td>#ICE</td><td>Avg↑</td><td>Std↓</td></tr><tr><td>5</td><td>5</td><td>38.4</td><td>2.1</td><td>5</td><td>40.4</td><td>1.2</td></tr><tr><td>10</td><td>10</td><td>41.3</td><td>3.6</td><td>5</td><td>40.8</td><td>2.4</td></tr><tr><td>20</td><td>20</td><td>42.6</td><td>3.3</td><td>5</td><td>42.2</td><td>2.0</td></tr><tr><td>30</td><td>30</td><td>40.0</td><td>2.9</td><td>5</td><td>41.1</td><td>1.9</td></tr></table>
+
+# 4.6 Ablation Studies
+
+We perform comprehensive ablation studies on our few-shot prompt including the impact of different video representation, number of shots and in-context selection. All the ablation results are evaluated on MSVD_QA validation set, and we report the mean and standard deviation of each setting on three sets of randomly sampled shots. For the cases with in-context example selection, we further select 5 examples as in-context examples from the sampled shots, while for the cases without in-context selection, all shots will be feed into the prompt. In Table 6, we show adding visual tokens consistently improves not only the model accuracy but also the model variance. A lower standard deviation indicates that the model is less sensitive to the few-shot sampling.
+
+To further demonstrate the impact of the additional temporal dimension of videos, we perform two ablations on the "Frame+Object+Event+Attribute" setting. First, we reduce the number of frame captions and visual tokens to be one9 for each video. We found that the performance drops significantly compared with using the default four frames, which indicates the model’s ability to incorporate information from multiple timestamps. Further, we found that fine-grained temporal modeling is rarely required for performing well on current video-language benchmarks. As shown in the ablation result where we reverse the order of all visual tokens and frame captions, the performance decreased only marginally, which indicates that current benchmarks may not be sufficient in reflecting the benefits from better temporal ordering.
+
+In Table 7, we first show that, with the same context length, namely, 5 in-context examples, in-context example selection significantly increases the performance as well as the robustness. At 10-shot, and 20-shot, directly fitting more shots into the prompt results in better performance. In-context selection achieves slightly lower performance but with significantly better efficiency due to shorter context. Interestingly, at 30-shot, in-context selection with 5 examples outperforms directly adding all 30 shots into the prompt. This is showing that in-context selection can help the model utilize a larger number noisy video examples. Nevertheless, we still observe that the benefit of adding more shots saturated at around 20 to 30 shots, even if with in-context selection. we view this as a remaining challenging on how to make language models benefit from longer contexts.
+
+# 5 Conclusions, Limitations and Future Work
+
+This paper proposes VidIL, a few-shot Video-language Learner via Image and Language models. It demonstrates the strong ability of large-scale language models on performing video-to-text tasks when frame features are provided as unified text representations using image-language models. We propose a temporal order aware prompt by decomposing videos into a hierarchical structure, which is able to plug in multiple levels of frame features, along with speech transcripts. Without pretraining on videos, our model outperforms vision-language models learned from large-scale video datasets on a variety of few-shot tasks, such as domain-specific captioning, question answering, and future event prediction. One limitation of using unified textual representation is that we might lose low-level visual features which can be essential for some specific tasks, such as fine-grained spatial visual question answering. We also observe that current video-language benchmarks rarely require explicit temporal tracking on the frames and visual tokens. Future work will focus on leveraging large-scale language models for learning script knowledge from long videos where temporal dynamics are better emphasized.
+
+# 6 Broader Impact
+
+An open-domain few-shot video-language learner has a wide range of beneficial applications for society, such as automatically detecting violent or mature content in videos and helping people with vision impairment understand videos. However, since the language model is pretrained on massive internet-scale text data, there might be unexpected output that can have potential negative impact on the society, such as bias against people of a certain gender, race or sexuality. Future work and dedicated collaboration from the community are needed to alleviate the potential negative societal impact of large language models.
+
+# Acknowledgements
+
+We thank the anonymous reviewers helpful suggestions. This research is based upon work supported in part by U.S. DARPA AIDA Program No. FA8750-18-2-0014 and U.S. DARPA KAIROS Program Nos. FA8750-19-2-1004. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.
+
+References   
+[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 3   
+[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. ArXiv preprint, abs/2204.14198, 2022. 3, 7, 8   
+[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3   
+[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 2425–2433. IEEE Computer Society, 2015. 7, 8   
+[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan, 2005. Association for Computational Linguistics. 6   
+[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 1, 3, 5   
+[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 190–200, Portland, Oregon, USA, 2011. Association for Computational Linguistics. 6   
+[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104–120. Springer, 2020. 2   
+[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 1931–1942. PMLR, 2021. 3   
+[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248–255. IEEE Computer Society, 2009. 4   
+[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. ArXiv preprint, abs/2106.11097, 2021. 3   
+[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6201–6210. IEEE, 2019. 3   
+[13] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot learners. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3816–3830, Online, 2021. Association for Computational Linguistics. 2   
+[14] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. AllenNLP: A deep semantic natural language processing platform. In Proceedings of Workshop for NLP Open Source Software (NLP-OSS), pages 1–6, Melbourne, Australia, 2018. Association for Computational Linguistics. 2
+
+[15] Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh, Lucy Vanderwende, Michel Galley, and Margaret Mitchell. Visual storytelling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1233–1239, San Diego, California, 2016. Association for Computational Linguistics. 2
+
+[16] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976–12985, 2021. 2
+
+[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 4904–4916. PMLR, 2021. 2
+
+[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 5583–5594. PMLR, 2021. 2
+
+[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32–73, 2017. 4
+
+[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956–1981, 2020. 4
+
+[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learningvia sparse sampling. In CVPR, 2021. 1, 3
+
+[22] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. TVQA: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1369–1379, Brussels, Belgium, 2018. Association for Computational Linguistics. 3
+
+[23] Jie Lei, Licheng Yu, Tamara Berg, and Mohit Bansal. What is more likely to happen next? videoand-language future event prediction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8769–8784, Online, 2020. Association for Computational Linguistics. 1, 6, 8
+
+[24] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online, 2020. Association for Computational Linguistics. 1
+
+[25] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In arxiv, 2021. 1, 3, 8, 9
+
+[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ArXiv preprint, abs/2201.12086, 2022. 2, 3, 6, 7, 8
+
+[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34, 2021. 2
+
+[28] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: Hierarchical encoder for Video+Language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046–2065, Online, 2020. Association for Computational Linguistics. 1, 3
+
+[29] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121–137. Springer, 2020. 2
+
+[30] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL-04), pages 605–612, Barcelona, Spain, 2004. 6
+
+[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6
+
+[32] Xudong Lin, Gedas Bertasius, Jue Wang, Shih-Fu Chang, Devi Parikh, and Lorenzo Torresani. Vx2text: End-to-end learning of video-based text generation from multimodal inputs. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7001–7011. IEEE, 2021. 3, 4
+
+[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853–13863, 2022. 3
+
+[34] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang. Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. arXiv preprint arXiv:2206.02082, 2022. 1
+
+[35] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13–23, 2019. 2
+
+[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. ArXiv preprint, abs/2002.06353, 2020. 1, 3, 6
+
+[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. ArXiv preprint, abs/2104.08860, 2021. 3
+
+[38] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9876–9886. IEEE, 2020. 3
+
+[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 2630–2640. IEEE, 2019. 2, 3
+
+[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. ArXiv preprint, abs/2203.02155, 2022. 2, 6
+
+[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 6
+
+[42] Mandela Patrick, Po-Yao Huang, Yuki Markus Asano, Florian Metze, Alexander G. Hauptmann, João F. Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3
+
+[43] Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao. Fewshot natural language generation for task-oriented dialog. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 172–182, Online, 2020. Association for Computational Linguistics. 1 [44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748–8763. PMLR, 2021. 2, 3, 4
+
+[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. ArXiv preprint, abs/1910.10683, 2019. 1
+
+[46] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1–67, 2020. 3
+
+[47] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992, Hong Kong, China, 2019. Association for Computational Linguistics. 4, 6, 8
+
+[48] Timo Schick and Hinrich Schütze. Exploiting cloze-questions for few-shot text classification and natural language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 255–269, Online, 2021. Association for Computational Linguistics. 2
+
+[49] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. ArXiv preprint, abs/2201.08264, 2022. 1
+
+[50] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: Plugging visual controls in text generation. ArXiv preprint, abs/2205.02655, 2022. 3
+
+[51] Derek Tam, Rakesh R. Menon, Mohit Bansal, Shashank Srivastava, and Colin Raffel. Improving and simplifying pattern exploiting training. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4980–4991, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2
+
+[52] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100–5111, Hong Kong, China, 2019. Association for Computational Linguistics. 2
+
+[53] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200–212, 2021. 3
+
+[54] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566–4575. IEEE Computer Society, 2015. 6
+
+[55] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-tosequence learning framework. ArXiv preprint, abs/2202.03052, 2022. 3
+
+[56] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation learning for text-video retrieval. ArXiv preprint, abs/2203.07111, 2022. 9
+
+[57] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A largescale, high-quality multilingual dataset for video-and-language research. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 4580–4590. IEEE, 2019. 1, 6
+
+[58] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. ArXiv preprint, abs/2108.10904, 2021. 2
+
+[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. ArXiv preprint, abs/2109.01652, 2021. 5
+
+[60] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017, pages 1645–1653, 2017. 1
+
+[61] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1 [62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288–5296. IEEE Computer Society, 2016. 1, 6 [63] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. ArXiv preprint, abs/2109.05014, 2021. 1, 3, 4 [64] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian, Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework. ArXiv preprint, abs/2205.01818, 2022. 1 [65] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3208–3216, 2021. 2 [66] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. ArXiv preprint, abs/2111.11432, 2021. 2 [67] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. ArXiv preprint, abs/2201.02639, 2022. 1 [68] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 8 [69] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. ArXiv preprint, abs/2204.00598, 2022. 3, 4 [70] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579–5588, 2021. 2 [71] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. ArXiv preprint, abs/2205.01068, 2022. 1 [72] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In Marina Meila and Tong Zhang, editors, Proceedings of the   
+38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume   
+139 of Proceedings of Machine Learning Research, pages 12697–12706. PMLR, 2021. 1, 6 [73] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pre-training for image captioning and VQA. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 13041–13049. AAAI Press, 2020. 2 [74] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In Sheila A. McIlraith and Kilian Q. Weinberger, editors, Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 7590–7598. AAAI Press, 2018. 3, 6 [75] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. ArXiv preprint, abs/2112.01522, 2021. 3

parse/dev/_LceCyuVcH/_LceCyuVcH_content_list.json

@@ -0,0 +1,1410 @@
+[
+    {
+        "type": "text",
+        "text": "Language Models with Image Descriptors are Strong Few-Shot Video-Language Learners ",
+        "text_level": 1,
+        "bbox": [
+            179,
+            122,
+            820,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhenhailong Wang1∗, Manling $\\mathbf { L i } ^ { 1 }$ ∗, Ruochen $\\mathbf { X } \\mathbf { u } ^ { 2 }$ , Luowei Zhou2†, Jie Lei3, Xudong Lin4, Shuohang $\\mathbf { W a n g } ^ { 2 }$ , Ziyi $\\mathbf { Y a n g } ^ { 2 }$ , Chenguang $\\mathbf { Z } \\mathbf { h } \\mathbf { u } ^ { 2 }$ , Derek Hoiem1, Shih-Fu Chang4, Mohit Bansal3, Heng Ji1 1UIUC 2MSR 3UNC 4Columbia University {wangz3,hengji}@illinois.edu ",
+        "bbox": [
+            251,
+            223,
+            750,
+            299
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract ",
+        "text_level": 1,
+        "bbox": [
+            462,
+            333,
+            535,
+            351
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The goal of this work is to build flexible video-language models that can generalize to various video-to-text tasks from few examples. Existing few-shot video-language learners focus exclusively on the encoder, resulting in the absence of a video-totext decoder to handle generative tasks. Video captioners have been pretrained on large-scale video-language datasets, but they rely heavily on finetuning and lack the ability to generate text for unseen tasks in a few-shot setting. We propose VidIL, a few-shot Video-language Learner via Image and Language models, which demonstrates strong performance on few-shot video-to-text tasks without the necessity of pretraining or finetuning on any video datasets. We use image-language models to translate the video content into frame captions, object, attribute, and event phrases, and compose them into a temporal-aware template. We then instruct a language model, with a prompt containing a few in-context examples, to generate a target output from the composed content. The flexibility of prompting allows the model to capture any form of text input, such as automatic speech recognition (ASR) transcripts. Our experiments demonstrate the power of language models in understanding videos on a wide variety of video-language tasks, including video captioning, video question answering, video caption retrieval, and video future event prediction. Especially, on video future event prediction, our few-shot model significantly outperforms state-of-the-art supervised models trained on large-scale video datasets. Code and processed data are publicly available for research purposes at https://github.com/MikeWangWZHL/VidIL. ",
+        "bbox": [
+            232,
+            364,
+            766,
+            655
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            680,
+            310,
+            698
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "One major gap between artificial intelligence and human intelligence lies in their abilities to generalize and perform well on new tasks with limited annotations. Recent advances in large-scale pre-trained generative language models [45, 6, 71, 24] have shown promising few-shot capabilities [72, 43, 63] in understanding natural language. However, few-shot video-language understanding is still in its infancy. A particular limitation of most recent video-language frameworks [28, 21, 61, 68, 67, 25, 64, 34] is that they are encoder-only, which means they do not have the ability to generate text from videos for purposes such as captioning [62, 57], question answering [60], and future prediction [23]. Meanwhile, unified video-language models [36, 49] that are capable of language decoding still rely heavily on finetuning using a large number of manually annotated video-text pairs, therefore cannot adapt quickly to unseen tasks. Few-shot video-to-text decoding is challenging because the natural language supervision for learning video-language representation is typically based on subtitles and automatic speech recognition (ASR) transcripts [39, 68], which differ significantly from downstream tasks in terms of distribution and may have poor semantic alignment across vision and text modalities. ",
+        "bbox": [
+            174,
+            712,
+            825,
+            864
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            90,
+            823,
+            119
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We propose to address this problem by harnessing the few-shot power of frozen large-scale language models, such as InstructGPT [40]. Our inspiration is derived from the fact that humans are excellent visual storytellers [15], with the ability to piece together a coherent story from a few isolated images. To mimic this, we propose VidIL, a few-shot Video-language Learner via Image and Language models, to use image models to provide information about the visual content in the video (as well as optionally use ASR to represent speech), and then we instruct language models to generate a video-based summary, answer, or other target output for diverse video-language tasks. ",
+        "bbox": [
+            173,
+            126,
+            825,
+            223
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The main challenge of understanding videos is that, videos contain rich semantics and temporal content at multiple granularities. Unlike static images which depict objects, attributes and events in a snapshot, the temporal dimension of videos further conveys the state changes of the objects, actions, and events. For example, in Figure 1, the individual frame captions of the video clip only describe static visual features such as \"a person holding a green object in hand\". In contrast, a correct video-level description would be \"a woman makes realistic looking leaves and flowers for a cake\", which involves reasoning over a collection of objects and events that occur at different timestamps in the video clip, such as \"cake decorating\" and \"flowered design\". Hence, to inform video-level description and queries, we need to represent all of this information and its temporal ordering. ",
+        "bbox": [
+            174,
+            231,
+            485,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/b5f930daf9032229a0d0a4e399381948a20cacaccc02903750ded88494988f0e.jpg",
+        "image_caption": [
+            "Figure 1: Multiple levels of information in videos. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            508,
+            252,
+            812,
+            448
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the unique challenges of videos, we propose to decompose a video into three levels: the video output, frame captions, and visual tokens (including objects, events, attributes). One major benefit from this hierarchical video representation is that we can separate the visual and temporal dimensions of a video. We leverage frozen image-language foundational models at lower levels to collect salient visual features from the sparsely sampled frames. Specifically, we first leverage a pretrained image-language contrastive model CLIP [44] to perform visual tokenization, based on the similarity score between frames and tokens of objects, events and attributes. The tokenization is done under the guidance of semantics role labeling [14], which provides us with candidate events with involved objects and related attributes. Next, in order to capture the overall semantics at the frame level, we employ the pretrained image captioner in the image-language model BLIP [26] to obtain frame captions. Finally, we instruct a pretrained large language model using in-context learning [40, 13, 51, 48] to interpret visual tokens and frame captions into the target textual output. In detail, we temporally order visual tokens and frame captions using specially designed prompts such as “First...Then...Finally”, to instruct the pretrained language model to track the changes of objects, events, attributes and frame semantics along the temporal dimension. ",
+        "bbox": [
+            174,
+            498,
+            825,
+            704
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Without pretraining or finetuning on any video datasets, we show that our approach outperforms both video-language and image-language state-of-the-art baselines on few-shot video captioning and question answering tasks. Moreover, on video-language event prediction, our approach significantly outperforms fully-supervised models while using only 10 labeled examples. We further demonstrate that our generative model can benefit broader video-language understanding tasks, such as text-video retrieval, via pseudo label generation. Additionally, we show that our model is highly flexible in adding new modalities, such as ASR transcripts. ",
+        "bbox": [
+            174,
+            712,
+            825,
+            808
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            827,
+            321,
+            843
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Image-Language Models and Their Applications on Video-Language Tasks ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            857,
+            733,
+            872
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Large-scale image-language pretraining models optimize image-text matching through contrastive learning [44, 17] and multimodal fusion [65, 27, 58, 66, 35, 52, 8, 29, 73, 70, 18, 16]. Recently, ",
+        "bbox": [
+            173,
+            883,
+            826,
+            911
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "BLIP [26] proposes a bootstrapping image-language pretraining framework with a captioner and a filterer which has shown promising performance on various image-language tasks. However, video-language pretraining [25, 36, 28, 38, 3, 1, 42, 33] is still hindered by noisy and domain-specific video datasets [74, 22, 39]. Naturally, researchers start to explore transferring the rich knowledge from image models to videos. Different from the traditional way of representing videos by 3D dense features [12], recent work [21, 25] proves that sparse sampling is an effective way to represent videos, which facilitates applying pre-trained image-language models to video-language tasks [37, 11]. Specifically, the image-language model BLIP [26] sets new state-of-the-art on zero-shot retrieval-style video-language tasks, such as video retrieval and video question answering. However, for generationstyle tasks such as domain-specific video captioning, video-language model UniVL [36] still leads the performance but highly rely on fine-tuning. In this work, we extend the idea of leveraging image-language models to a wide variety of video-to-text generation tasks. We further connect imagelanguage models with language models which empowers our model with strong generalization ability. We show that the knowledge from both image-language pretraining and language-only pretraining can benefit video-language understanding in various aspects. ",
+        "bbox": [
+            174,
+            92,
+            825,
+            299
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Unifying MultiModal Tasks with Language Models ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            325,
+            568,
+            342
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The community has paid much attention to connecting different modalities with a unified representation recently. Text-only generation models, such as T5 [46], have been extended to vision-language tasks by text generation conditioned on visual features [9, 53, 50, 75, 55]. In order to fully leverage the generalization power from pretained language models, [63] represents images using text in a fully symbolic way. [32] includes more modalities such as video and audio, but requires annotated video-text data to jointly training the language model with the video and audio tokenizer. In this work, we propose a temporal-aware hierarchical representation for describing a video textually. To our knowledge, we are the first work to leverage prompting a frozen language model for tackling few-shot video-language tasks with a unified textual representation. Concurrent work Socratic [69] uses a zero-shot language-based world-state history to represent long videos with given time stamps, while our model can quickly adapt to different video and text distributions with few examples. Furthermore, we show that by injecting temporal markers to the prompt we can make a pre-trained language model understand fine-grained temporal dynamics in video events. Compared with the concurrent work Flamingo [2], which requires dedicated vision-language post-pretraining, our framework does not require to pretrain or finetune on any video data. Our framework is simple and highly modulated where all the components are publicly available. Additionally, our framework is more flexible on adding new modalities, e.g., automatic speech recognition, without the need for complex redesigning. ",
+        "bbox": [
+            174,
+            356,
+            825,
+            592
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            621,
+            269,
+            638
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We propose a hierarchical video representation framework which decomposes a video into three levels, i.e., visual token level, frame level and video level. The motivation is to separate the spatial and temporal dimension of a video in order to leverage image-language and language-only foundation models, such as CLIP [44] and GPT-3 [6]. All three levels use a unified textual representation which enables us to leverage the powerful few-shot ability from pretrained language models. ",
+        "bbox": [
+            174,
+            660,
+            825,
+            729
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Frame Level: Image Captioning ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            756,
+            436,
+            772
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following [21] we first perform sparse sampling to obtain several video frames. Unless otherwise specified, we sample 4 frames for frame level and 8 frames for visual token level. We then feed each frame into a pre-trained image-language model to obtain frame level captions. An example can be found in the blue part of Figure 2. In our experiments, we use BLIP [26], a recent image-language framework containing both image-grounded encoder and decoder, for generating frame captions. We follow [26] to do both captioning and filtering on each frame. However, as mentioned in Section 1, videos contain rich semantics and temporal contents at multiple granularities. It is not enough to generate video-level target text such as video captions solely based on frame captions. Thus, we further perform visual tokenization for each frame to capture features at a finer granularity. ",
+        "bbox": [
+            174,
+            786,
+            825,
+            911
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/c144de0cb595e241110b349a7da85b61029ef363498acc4c08b0c7647b08efae.jpg",
+        "image_caption": [
+            "Figure 2: Overview of VidIL framework. We represent a video in a unified textural representation containing three semantic levels: visual token level, frame level, and video level. At visual token level, we extract salient objects, events, attributes for each sampled frame. At frame level, we perform image captioning and filtering. At video level, we construct video representation by aggregating the visual tokens, frame captions and other text modalities such as ASR, using a few-shot temporalaware prompt. We then feed the prompt to a pre-trained language model together with task-specific instructions to generate target text for a variety of video-language tasks. Examples of the full prompt for different tasks can be found in Appendix ??. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            179,
+            88,
+            816,
+            411
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Visual Token Level: Structure-Aware Visual Tokenization ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            561,
+            612,
+            577
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "At this level, we aim to extract the textual representations of salient visual token types, such as objects, events and attributes. We found that pre-defined classes for classification, such as those in ImageNet [10], are far from enough for covering the rich semantics in open-domain videos. Thus, instead of using classification-based methods for visual tokenization as in previous work [32, 63], we adopt a retrieval-based visual tokenization approach by leveraging pre-trained contrastive imagelanguage models. Given a visual token vocabulary which contains all candidate object, event, and attribute text phrases, we compute the image embedding of a frame and the text embeddings of the candidate visual tokens using a contrastive multi-modal encoder, CLIP [44]. We then select top 5 visual tokens per frame based on the cosine similarity of the image and text embeddings. An example of the extracted object tokens can be found in the green part of Figure 2. ",
+        "bbox": [
+            174,
+            588,
+            825,
+            727
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Unlike in images where objects and attributes already cover most visual features, events are more informative in videos. In order to discover events from video frames, we construct our own event vocabulary by extracting event structures from Visual Genome [19] synsets3 using Semantic Role Labeling. Specifically, we first select the phrases that contain at least one verb and one argument as events. Then we remove highly similar events based on their sentence similarity using SentenceBERT [47] embeddings. For object vocabulary, we adopt OpenImage [20] full classes $( { \\sim } 2 0 \\mathbf { k } )$ , instead of using the visually groundable subset $( \\sim 6 0 0 )$ as in concurrent work [69]. We found that using large but noisy vocabulary is more effective than using small but clean vocabulary in our retrieval-based setting with CLIP. For attribute vocabulary, we adopt visual genome attribute synset. In Section 4.6, we provide ablation study on the impact of different types of visual tokens. The statistics of visual token vocabulary can be found in Appendix Table ??. ",
+        "bbox": [
+            174,
+            733,
+            825,
+            885
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/331592c36e0944e6b5f6973236f3089e33d229d668d68f0224a0f67fdb784fba.jpg",
+        "image_caption": [
+            "Figure 3: Temporal-aware prompt successfully distinguishes the Sunset and Sunrise scenarios based on the temporal ordering change of objects and frame captions, while the static prompt fails. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            176,
+            93,
+            782,
+            231
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Video Level: Temporal-Aware Few-shot Prompting ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            284,
+            568,
+            299
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Once we obtain the textual representation from frame level and visual token level, the final step is to put the pieces together to generate a video level target text. The goal is to build a model that can be quickly adapted to any video-to-text generation task with only a few examples. To this end, we propose to leverage large-scale pre-trained language models, such as GPT-3 [6], with a temporal-aware few-shot prompt. As shown in Figure 2, our framework can be readily applied to various video-to-text generation tasks, such as video captioning and video question answering, with a shared prompt template. The proposed prompting strategy enables a language model to attend to the lower level visual information as well as taking into account the temporal ordering. ",
+        "bbox": [
+            173,
+            309,
+            825,
+            421
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here, we use the video captioning task depicted in Figure 2 to illustrate the details. The few-shot prompt consists of three parts: instruction, few-shot context, and task query. The instruction is a concise description of the generation task, e.g., \"Generate a video caption based on the objects, events, attributes and frame captions. Example:\", which is proved to be effective in zero-shot and few-shot settings [6, 59]. The few-shot context contains the selected in-context examples as well as the test video instance. Each video instance is represented by the aggregated visual tokens4, e.g., \"Objects: First, bath toy. Then,...\" the frame captions, such as \"Frame Captions: First, a toddler playing in a bathtub filled with toys. Then,...\", and the ASR inputs if available, e.g., \"Subtitle:<ASR Transcript>\". Finally, the task query is a task-specific suffix indicating the target text format, e.g. \"Video Caption:\". For in-context examples (omitted here for simplicity), the task query is followed by ground truth annotation, while for the test instance, the generation starts at the end of the task query. ",
+        "bbox": [
+            174,
+            428,
+            826,
+            607
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Formally, we denote the instruction line as $\\mathbf { t }$ , few-shot context as c, the task query as q, and the target text as $\\mathbf { y }$ , where $\\mathbf { y } = ( y _ { 1 } , y _ { 2 } , . . . , y _ { L } )$ . The generation of the next target token $y _ { l }$ can be modeled as: ",
+        "bbox": [
+            173,
+            613,
+            823,
+            642
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "img_path": "images/bc7c063980fb4da70b5248c3574a151aaa1f1ceeacd34ba3434960d3c7181041.jpg",
+        "text": "$$\ny _ { l } = \\mathop { \\arg \\operatorname* { m a x } } _ { y } p ( y | \\mathbf { s } , \\mathbf { c } , \\mathbf { q } , y _ { < l } )\n$$",
+        "text_format": "latex",
+        "bbox": [
+            397,
+            657,
+            599,
+            684
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In order to capture the temporal dynamics between frames and visual tokens, we further propose to inject temporal markers to the prompt. As shown in the few-shot context in Figure 2, each visual token and frame caption is prefixed with a natural language phrase indicating its temporal ordering, e.g., \"First,\",\"Then,\", and \"Finally,\". We found adding the temporal marker can make the language model conditioned on not only literal but also temporal information of the context. We show an example in Figure 3, where we compare our temporal-aware prompt with a static prompt on video captioning using InstructGPT. Again, the in-context examples are omitted here, which can be found in Appendix ??. In this example, the only difference between these two contexts is the ordering of the visual tokens and the frame captions. For the context on the left, where \"sun moving\" appears before \"night sky\", we are expected to see a story talking about sunset, while for the context on the right, we are expected to see sunrise. We can see the static prompt generates captions about sunset for both contexts, while the temporal-aware prompt can capture temporal ordering correctly and generate sunrise for the context on the right. ",
+        "bbox": [
+            173,
+            690,
+            825,
+            871
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            89,
+            312,
+            107
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setup ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            121,
+            351,
+            136
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To comprehensively evaluate our model, we show results on four video-language understanding tasks in few-shot settings: video captioning, video question answering (QA), video-language event prediction, and text-video retrieval. We compare our approach with state-of-the-art approaches on five benchmarks, i.e, MSR-VTT [62], MSVD [7], VaTeX [57], YouCook2 [74], and VLEP [23]. The statistics of the datasets can be found in Table 1. For more details please refer to Appendix ??. ",
+        "bbox": [
+            174,
+            147,
+            825,
+            215
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. We use CLIP-L/146 as our default encoder for visual tokenization. We adopt BLIP captioning checkpoint7 finetuned on COCO [31] for frame captioning. We use InstructGPT [40] as our default language model for generating text conditioned on the few-shot prompt. To construct event vocabulary, we use the semantic role labeling model from AllenNLP8. The experiments are conducted on 2 NVIDIA V100 (16GB) GPUs. All few-shot ",
+        "bbox": [
+            174,
+            226,
+            485,
+            363
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/92dd6fadc5a80d33d4e99b29ba373e80a80afa39fcabf656f24d4c71b6eab48f.jpg",
+        "table_caption": [
+            "Table 1: Statistics of datasets in our experiments "
+        ],
+        "table_footnote": [
+            "finetuning on baselines and semi-supervised training are performed on 2 Nvidia V100 16G GPUs. "
+        ],
+        "table_body": "<table><tr><td>Dataset</td><td>Task</td><td>Split Count # train /#eval</td></tr><tr><td>MSR-VTT[62]</td><td>Captioning; QA</td><td>6,513 /2.990</td></tr><tr><td>MSR-VTT[62]</td><td>Retrieval</td><td>7,010 / 1,000</td></tr><tr><td>MSVD [7]</td><td>Question Answering</td><td>30,933 /13,157</td></tr><tr><td>VaTeX v1.15 [57]</td><td>Captioning; Retrieval</td><td>25,991/6.000</td></tr><tr><td>YouCook2[74]</td><td>Captioning</td><td>10,337 /3,492</td></tr><tr><td>VLEP [23]</td><td>Event Prediction</td><td>20,142/4,192</td></tr></table>",
+        "bbox": [
+            496,
+            246,
+            823,
+            352
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In-context Example Selection. From our preliminary experiments, we find that the generation performance is sensitive to the quality of in-context examples. For example, for QA tasks such as MSVD-QA where the annotations are automatically generated, the <question, answer> pair in randomly selected in-context examples can be only weakly-correlated with the video context. Thus, instead of using a fixed prompt for each query, we dynamically filter out the irrelevant in-context examples. Specifically, given a randomly sampled $M .$ -shot support set from the training set, we select a subset of $N _ { ☉ }$ -shots as in-context examples based on their SentenceBERT [47] similarities with text queries. Furthermore, we reorder the selected examples in ascending order based on the similarity score to account for the recency bias [72] in large language models. For QA tasks, we choose the most relevant in-context examples by comparing with questions. While for captioning task, we compare with frame captions. If not otherwise specified, we use $M { = } I O$ and $N { = } 5$ , which we consider as 10-shot training. ",
+        "bbox": [
+            173,
+            395,
+            825,
+            560
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Few-shot Video Captioning ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            577,
+            401,
+            592
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We report BLEU-4 [41], ROUGE-L [30], METEOR [5], and CIDEr [54] scores on three video captioning benchmarks covering both open-domain (MSR-VTT, VaTeX) and domain-specific (YouCook2) videos. We compare with both state-of-the-art video captioner (UniVL [36]) and image captioner (BLIP [26]). In order to implement the BLIP baseline for few-shot video captioning, we extend the approach used for text-video retrieval evaluation in [26] to video-language training. Specifically, we concatenate the visual features of sampled frames and then feed them into the image-grounded text-encoder to compute the language modeling loss. This is equivalent to stitching the sampled frames into a large image and then feeding it to BLIP for image captioning. We found that this simple approach results in very strong baselines. ",
+        "bbox": [
+            174,
+            603,
+            825,
+            728
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 2, existing methods have strong bias on certain datasets. For example, UniVL performs well on YouCook2 but fails on MSR-VTT and VaTeX, while BLIP performs the opposite. This is because UniVL is pretrained on HowTo100M which favors instructional videos, i.e., YouCook2, while BLIP is pre-training on image-caption pairs which favors description-style captions, i.e., MSR-VTT and VaTeX. On the contrary, our model performs competitively on both open-domain and instructional videos, and significantly outperforms the baselines on the average CIDEr score across all three benchmarks. This indicates that by leveraging language models, we can maintain strong few-shot ability regardless of the video domain or the target caption distribution. ",
+        "bbox": [
+            173,
+            734,
+            825,
+            845
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/bdc0755cadf49e276175a4fa571481c0553518341285949af409281797b235b2.jpg",
+        "table_caption": [
+            "Table 2: 10-shot video captioning results. ♠ indicates concurrent work. The reported Flamingo [2] results are using 16 shots. #VideoPT represents the number of videos used for pre-training. B-4, R-L, $M$ , $C$ represents BLEU-4, ROUGE-L, METEOR and CIDEr. Avg $C$ represents the average CIDEr score across all available benchmarks. ASR indicates whether the model has access to the ASR subtitles. $B L I P$ and $B L I P _ { c a p }$ use the pretrained checkpoint and the finetuned checkpoint on COCO captioning. All results are averaged over three random seeds. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">#VideOpr ASR</td><td rowspan=\"2\"></td><td rowspan=\"2\">B-4R-LM</td><td colspan=\"3\">MSR-VTTCaption</td><td colspan=\"3\">YouCook2 Caption C</td><td rowspan=\"2\"></td><td colspan=\"3\">VaTex Caption</td><td rowspan=\"2\">Avg C</td></tr><tr><td></td><td></td><td>C</td><td>B-4R-LM</td><td></td><td></td><td>B-4 R-L M</td><td></td><td>C</td></tr><tr><td>Few-shot</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>UniVL</td><td>1.2M</td><td>No</td><td>2.1 22.5 9.5</td><td></td><td></td><td>3.6</td><td>3.3</td><td>25.3 11.6</td><td>34.1</td><td>1.7</td><td>15.7</td><td>8.0</td><td>2.1</td><td>13.3</td></tr><tr><td>BLIP</td><td>0</td><td>No</td><td>27.7 43.0 23.0 39.5</td><td></td><td></td><td></td><td>0.7</td><td>9.0 3.4</td><td>11.5</td><td>13.5</td><td>39.5</td><td>15.4</td><td>20.7</td><td>23.9</td></tr><tr><td>BLIPcap</td><td>0</td><td>No</td><td>21.648.022.7 30.2</td><td></td><td></td><td></td><td>3.7</td><td>8.6 3.8</td><td>9.4</td><td>20.7</td><td>41.5</td><td>17.4</td><td>28.9</td><td>22.8</td></tr><tr><td>VidIL(ours)</td><td>0</td><td>No</td><td>26.0 51.7 24.7 36.3</td><td></td><td></td><td></td><td>2.6</td><td>22.9 9.5</td><td>27.0</td><td>22.2</td><td></td><td>43.620.0</td><td>36.7</td><td>33.3</td></tr><tr><td>UniVL</td><td>1.2M</td><td>Yes</td><td>-</td><td></td><td>=</td><td>-</td><td>4.3</td><td>26.4 12.2</td><td>48.6</td><td>2.7</td><td>17.7</td><td>10.2</td><td>3.4</td><td>26.0</td></tr><tr><td>VidIL(ours)</td><td>0</td><td>Yes</td><td></td><td></td><td></td><td></td><td>10.7 35.9</td><td></td><td>19.4 111.6</td><td></td><td>23.2 44.2 20.6 38.9</td><td></td><td></td><td>75.3</td></tr><tr><td>Flamingo-3B(16)</td><td>27M</td><td>No</td><td></td><td></td><td></td><td></td><td></td><td></td><td>73.2</td><td></td><td></td><td></td><td>57.1</td><td></td></tr><tr><td>Flamingo-80B(16) 27M</td><td></td><td>No</td><td></td><td></td><td></td><td></td><td></td><td></td><td>84.2</td><td></td><td></td><td></td><td>62.8</td><td>=</td></tr><tr><td>Fine-tuning</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>UniVL</td><td>1.2M</td><td>No</td><td>42.0 61.0 29.0 50.1</td><td></td><td></td><td></td><td></td><td></td><td></td><td>[11.2 40.1 17.6 127.0|22.8 38.6 22.3 33.4</td><td></td><td></td><td></td><td>70.2</td></tr><tr><td>UniVL</td><td>1.2M</td><td>Yes</td><td>=</td><td></td><td></td><td></td><td></td><td></td><td></td><td>16.6 45.7 21.6176.823.7 39.3 22.7 35.6</td><td></td><td></td><td></td><td>106.2</td></tr></table>",
+        "bbox": [
+            173,
+            189,
+            825,
+            393
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As discussed in Section 1, video captions describe the content in various semantic levels. The N-gram based metric may not fairly reflect the models’ performance in capturing the video-caption alignment. We further verify this hypothesis in Section 4.5. Thus, in addition to automatic metrics, we include qualitative examples illustrated in Figure 4. More examples are in Appendix ??. ",
+        "bbox": [
+            173,
+            420,
+            825,
+            476
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Additionally, for most existing methods and also concurrent work, e.g., Flamingo [2], adding a new modality often requires a dedicated model redesign or retraining. However, the nature of our framework, where we use a unified textual representation for each level, makes it highly flexible for incorporating new modalities. As shown in row 6 in Table, our model can effectively utilize extra information from ASR to obtain significantly better few-shot performance on certain datasets such as YouCook2. ",
+        "bbox": [
+            173,
+            482,
+            825,
+            565
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/42488e5e146be037d7c57abcc48a78f8af1cf75e6517bc50d3e1b32bfd715b6a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            173,
+            575,
+            823,
+            756
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: Qualitative examples on video captioning. Grey boxes contain part of the video representation from our model. Blue boxes contain caption generation from different models. Green boxes contain ground truth annotations. Bold green text highlights the correct information that is not captured in baseline outputs which can be reasoned from our visual tokens and frame captions. ",
+        "bbox": [
+            173,
+            768,
+            826,
+            825
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Few-shot Video Question Answering ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            843,
+            465,
+            859
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare the test accuracy of our approach with few-shot pretrained BLIP, $\\mathsf { B L I P } _ { V Q A }$ [26], and concurrent work Flamingo [2] on two video question answering benchmarks, MSR-VTT_QA and MSVD_QA. $\\mathsf { B L I P } _ { V Q A }$ represents finetuned BLIP on VQA [4] dataset, which is the previous SOTA on zero/few-shot video question answering. In order to have fairer comparison with $\\mathsf { B L I P } _ { V Q A }$ , we reduce the shot number to 5 and report the average accuracy on three sets of randomly selected 5-shot examples. As shown in Table 3, our method outperforms previous SOTA by a large margin. Comparing with concurrent work Flamingo, which is post-pretrained on a large number of video-text data, our model is training-free and did not observe any video data. However, with only imagelanguage and language-only knowledge, our 5-shot model is able to outperform 8-shot Flamingo-3B and achieve on-par performance with 4-shot Flamingo-80B. ",
+        "bbox": [
+            174,
+            869,
+            825,
+            911
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            90,
+            826,
+            189
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/80c20213acc3c9f54437b9d3f7c1181c8608963e7d4b561375c9c1f29b5cc5cb.jpg",
+        "table_caption": [
+            "Table 3: Video QA results. $\\mathsf { B L I P } _ { V Q A }$ is finetuned on VQA [4]. ♠ indicates concurrent work. PT, FT indicates pretraining and finetuning. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>#videoPT</td><td>#videoFT</td><td>MSR-VTT</td><td>MSVD</td></tr><tr><td>BLIP</td><td>0</td><td>O-shot</td><td>0.55</td><td>0.45</td></tr><tr><td>BLIP</td><td>0</td><td>5-shot</td><td>0.84</td><td>0.53</td></tr><tr><td>BLIPvQA [26]</td><td>0</td><td>O-shot</td><td>19.2</td><td>35.2</td></tr><tr><td>VidIL(ours)</td><td>0</td><td>5-shot</td><td>21.2</td><td>39.1</td></tr><tr><td>Flamingo-3B [2]</td><td>27M</td><td>4-shot</td><td>14.9</td><td>33.0</td></tr><tr><td>Flamingo-3B_[2]</td><td>27M</td><td>8-shot</td><td>19.6</td><td>37.0</td></tr><tr><td>Flamingo-80B [2]</td><td>27M</td><td>4-shot</td><td>23.9</td><td>41.7</td></tr><tr><td>Flamingo-80B [2]</td><td>27M</td><td>8-shot</td><td>27.6</td><td>45.5</td></tr><tr><td>ALPRO [25]</td><td>2M</td><td>full-shot</td><td>42.1</td><td>45.9</td></tr></table>",
+        "bbox": [
+            176,
+            238,
+            614,
+            386
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/73514530509fe457864a28260c85bbc7f3012661dc1bd0d43e115e6c87eefb5c.jpg",
+        "table_caption": [
+            "Table 4: Accuracy $( \\% )$ on VLEP hidden test set. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>#videoFT</td><td>Acc</td></tr><tr><td>VLEP [23] MERLOT[68]</td><td>20142 20142</td><td>67.5 68.4</td></tr><tr><td>VidIL(ours)</td><td>10-shot</td><td>72.0</td></tr><tr><td>Human</td><td>-</td><td>90.5</td></tr></table>",
+        "bbox": [
+            625,
+            239,
+            826,
+            325
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4 Few-shot Video-Language Event Prediction ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            404,
+            514,
+            419
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we show that our model not only can answer questions about the video visual features but also answering \"What is more likely to happen next?\". Given a video with associated subtitle transcript as premise, the video-language event prediction (VLEP) task is to predict the most likely future event. The original VLEP [23] paper formulates the problem as a binary classification problem where the model will be chosen from two possible future event candidates. Instead, we formulate this problem as another video-to-text generation problem to fit into our framework. Figure 5 depicts an example with the same format as in Figure 2. Similar to the evaluation setting in QA, the generated free-form text will first be mapped to one of the two candidate answers using SentenceBert [47], and then calculate the accuracy. In Table 4, we report accuracy on the hidden test set of VLEP [23]. To our surprise, our 10-shot model outperforms state-of-the-art fully-supervised baseline, i.e., MERLOT [68], by a large margin $( \\sim 4 \\% )$ . This shows that our model has strong few-shot ability not only on videolanguage understanding but also on prediction. Since event prediction tasks rely heavily on temporal ordering, we show that with the proposed temporal-aware prompting, language models can be guided to capture temporal dynamics between historical and future events. ",
+        "bbox": [
+            173,
+            429,
+            825,
+            622
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/337717ccf8591fcb87cdc2bcb90b9982ea93bafdcb263be40f57ceb4066b57ea.jpg",
+        "image_caption": [
+            "Figure 5: Prompt for VLEP task. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            187,
+            636,
+            805,
+            737
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Semi-supervised Text-Video Retrieval ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            775,
+            475,
+            790
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In addition to video-to-text generation tasks, we show that a broader range of video-language tasks can benefit from our few-shot video captioner from a data perspective. Here, we consider a lowbudget semi-supervised setting where we only have a few labeled video-caption pairs and a large amount of unlabeled videos. The idea is to leverage our video captioner to generate pseudo labels for training any given vision-language models. As a case study, we evaluate on two text-video retrieval benchmarks, i.e., MSR-VTT and VaTeX. We use greedy decoding to generate pseudo caption for each video in the training set. We then train an identical base model, i.e., BLIP, using different pseudo labeled data as well as ground truth annotations. We report Recall $@$ 1 and 5 for both video-to-text and text-to-video retrieval. Table 5 shows that through training on our pseudo labels, we can achieve significant improvements compared with zero-shot BLIP. We also show that the performance gain is not simply a result of training on more data, since finetuning on the pseudo labels generated by other baselines (UniVL, BLIP) is less effective and can even hurt the performance. Furthermore, on MSR-VTT Recall $@$ 5 we can even achieve comparable performance against BLIP model finetuned on full ground truth annotations. ",
+        "bbox": [
+            173,
+            800,
+            825,
+            911
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3d91cbc690fa88f5cb9c62c6176d417056eb1bb8034ac54fa9a50416e7a1be80.jpg",
+        "table_caption": [
+            "Table 5: Semi-supervised text-video retrieval with 10 labeled examples. $\\mathrm { V _ { l a b e l } }$ or $\\mathrm { \\Delta V _ { u n l a b e l } }$ are the number of labeled and unlabeled videos, respectively. $t \\_ R I$ and $t \\_ R$ denote video-to-text Recall $@ 1$ and 5. $\\nu \\_ R I$ and $\\nu \\_ R 5$ denote text-to-video Recall $@ 1$ and 5. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Pseudo Label</td><td colspan=\"5\">MSR-VTTRetrieval</td><td colspan=\"5\">VaTexRetrieval</td></tr><tr><td>Vlabel/Vunlabel t_R1t_R5v_R1</td><td></td><td></td><td></td><td>v_R5</td><td>Vlabel/Vunlabel t_R1t_R5v_R1</td><td></td><td></td><td></td><td>v_R5</td></tr><tr><td>BLIP</td><td></td><td></td><td>33.2</td><td>57.2</td><td>40.5</td><td>62.8</td><td></td><td>28.2</td><td>53.4</td><td>34.0</td><td>58.6</td></tr><tr><td>BLIP</td><td>UniVL</td><td>10 /7010</td><td>33.1</td><td>57.3</td><td>33.6</td><td>57.7</td><td>10 /22685</td><td>25.5</td><td>47.7</td><td>26.1</td><td>49.1</td></tr><tr><td>BLIP</td><td>BLIP</td><td>10/7010</td><td>35.6</td><td>60.8</td><td>39.8</td><td>60.4</td><td>10/22685</td><td>26.3</td><td>50.5</td><td>29.3</td><td>53.6</td></tr><tr><td>BLIP</td><td>BLIPcap</td><td>10/7010</td><td>35.3</td><td>58.0</td><td>39.1</td><td>63.3</td><td>10/22685</td><td>23.9</td><td>46.8</td><td>27.5</td><td>49.7</td></tr><tr><td>BLIP</td><td>VidIL(ours)</td><td>10/7010</td><td>39.6</td><td>64.5</td><td>40.8</td><td>65.2</td><td>10 /22685</td><td>33.3</td><td>59.1</td><td>33.7</td><td>59.5</td></tr><tr><td>BLIP</td><td>Ground Truth</td><td>7010/0</td><td>43.6</td><td>66.2</td><td>43.1</td><td>67.2</td><td>22685/0</td><td>40.1</td><td>66.4</td><td>40.1</td><td>66.6</td></tr><tr><td>ALPRO [25]</td><td>Ground Truth</td><td>140200/0</td><td>32.0</td><td>60.6</td><td>33.9</td><td>60.7</td><td></td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DRL [56]</td><td>Ground Truth</td><td>180000/0</td><td>54.1</td><td>77.4</td><td>52.9</td><td>78.5</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr></table>",
+        "bbox": [
+            173,
+            138,
+            825,
+            268
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            291,
+            825,
+            375
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Another interesting observation is that, compared with the video captioning results in Table 2, we found that the gain of our model over baselines on text-video retrieval is more visible than on captioning. A key factor in performing well on text-video retrieval tasks is to learn a good video-text multi-modal alignment. This result shows that our pseudo labels capture richer video-text alignment that can benefit the retrieval-style downstream task. The N-gram based generation metrics, e.g., BLEU, may not be able to fully reflect the alignment information, due to the variety of semantic levels in video captions. Furthermore, from a data perspective, our video captioner can be viewed as a data augmentation tool which is capable of generating or augmenting any open-domain videolanguage pretraining datasets with minimal human effort. As a result, we can potentially improve video-language pretraining by constructing a cleaner and more diverse video-text corpus. ",
+        "bbox": [
+            173,
+            381,
+            826,
+            520
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/1216fec2ba1e8eca8e72af10e7c75e06a1a18c33372a4215061d9f781eb347cd.jpg",
+        "table_caption": [
+            "Table 6: Impact of visual tokens and temporal dimension. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Video Representation</td><td>Avg↑ Std↓</td><td></td></tr><tr><td rowspan=\"3\">Visual Token</td><td>Frame Frame+Object</td><td>39.6 40.3</td><td>3.7 2.9</td></tr><tr><td>Frame+Object+Event</td><td>39.9</td><td>2.8</td></tr><tr><td>Frame+Object+Attibute Frame+Object+Event+Attribute</td><td>40.9 40.8</td><td>2.9 2.4</td></tr><tr><td>Temporal</td><td></td><td>Reduce to one frame Reverse temporal order</td><td>38.5 40.7</td><td>2.4 1.7</td></tr></table>",
+        "bbox": [
+            178,
+            575,
+            516,
+            699
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/b6823d5dd736e75f8f7e9c3d796ad639cd35817997264d469a99028f8ff5a52a.jpg",
+        "table_caption": [
+            "Table 7: Impact of shot selection. #ICE indicates the number of in-context examples in the prompt. Details of in-context example selection are in the Appendix. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>#shot</td><td colspan=\"3\">w/o selection</td><td colspan=\"3\">w/ selection</td></tr><tr><td></td><td>#ICE</td><td>Avg↑</td><td>Std</td><td>#ICE</td><td>Avg↑</td><td>Std↓</td></tr><tr><td>5</td><td>5</td><td>38.4</td><td>2.1</td><td>5</td><td>40.4</td><td>1.2</td></tr><tr><td>10</td><td>10</td><td>41.3</td><td>3.6</td><td>5</td><td>40.8</td><td>2.4</td></tr><tr><td>20</td><td>20</td><td>42.6</td><td>3.3</td><td>5</td><td>42.2</td><td>2.0</td></tr><tr><td>30</td><td>30</td><td>40.0</td><td>2.9</td><td>5</td><td>41.1</td><td>1.9</td></tr></table>",
+        "bbox": [
+            527,
+            607,
+            825,
+            698
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4.6 Ablation Studies ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            737,
+            328,
+            752
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We perform comprehensive ablation studies on our few-shot prompt including the impact of different video representation, number of shots and in-context selection. All the ablation results are evaluated on MSVD_QA validation set, and we report the mean and standard deviation of each setting on three sets of randomly sampled shots. For the cases with in-context example selection, we further select 5 examples as in-context examples from the sampled shots, while for the cases without in-context selection, all shots will be feed into the prompt. In Table 6, we show adding visual tokens consistently improves not only the model accuracy but also the model variance. A lower standard deviation indicates that the model is less sensitive to the few-shot sampling. ",
+        "bbox": [
+            173,
+            765,
+            825,
+            877
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To further demonstrate the impact of the additional temporal dimension of videos, we perform two ablations on the \"Frame+Object+Event+Attribute\" setting. First, we reduce the number of frame captions and visual tokens to be one9 for each video. We found that the performance drops significantly compared with using the default four frames, which indicates the model’s ability to incorporate information from multiple timestamps. Further, we found that fine-grained temporal modeling is rarely required for performing well on current video-language benchmarks. As shown in the ablation result where we reverse the order of all visual tokens and frame captions, the performance decreased only marginally, which indicates that current benchmarks may not be sufficient in reflecting the benefits from better temporal ordering. ",
+        "bbox": [
+            173,
+            883,
+            821,
+            911
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            90,
+            825,
+            188
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "In Table 7, we first show that, with the same context length, namely, 5 in-context examples, in-context example selection significantly increases the performance as well as the robustness. At 10-shot, and 20-shot, directly fitting more shots into the prompt results in better performance. In-context selection achieves slightly lower performance but with significantly better efficiency due to shorter context. Interestingly, at 30-shot, in-context selection with 5 examples outperforms directly adding all 30 shots into the prompt. This is showing that in-context selection can help the model utilize a larger number noisy video examples. Nevertheless, we still observe that the benefit of adding more shots saturated at around 20 to 30 shots, even if with in-context selection. we view this as a remaining challenging on how to make language models benefit from longer contexts. ",
+        "bbox": [
+            174,
+            194,
+            825,
+            319
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusions, Limitations and Future Work ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            347,
+            563,
+            364
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This paper proposes VidIL, a few-shot Video-language Learner via Image and Language models. It demonstrates the strong ability of large-scale language models on performing video-to-text tasks when frame features are provided as unified text representations using image-language models. We propose a temporal order aware prompt by decomposing videos into a hierarchical structure, which is able to plug in multiple levels of frame features, along with speech transcripts. Without pretraining on videos, our model outperforms vision-language models learned from large-scale video datasets on a variety of few-shot tasks, such as domain-specific captioning, question answering, and future event prediction. One limitation of using unified textual representation is that we might lose low-level visual features which can be essential for some specific tasks, such as fine-grained spatial visual question answering. We also observe that current video-language benchmarks rarely require explicit temporal tracking on the frames and visual tokens. Future work will focus on leveraging large-scale language models for learning script knowledge from long videos where temporal dynamics are better emphasized. ",
+        "bbox": [
+            174,
+            383,
+            825,
+            563
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "6 Broader Impact ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            590,
+            338,
+            608
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "An open-domain few-shot video-language learner has a wide range of beneficial applications for society, such as automatically detecting violent or mature content in videos and helping people with vision impairment understand videos. However, since the language model is pretrained on massive internet-scale text data, there might be unexpected output that can have potential negative impact on the society, such as bias against people of a certain gender, race or sexuality. Future work and dedicated collaboration from the community are needed to alleviate the potential negative societal impact of large language models. ",
+        "bbox": [
+            174,
+            627,
+            825,
+            724
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            752,
+            338,
+            770
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We thank the anonymous reviewers helpful suggestions. This research is based upon work supported in part by U.S. DARPA AIDA Program No. FA8750-18-2-0014 and U.S. DARPA KAIROS Program Nos. FA8750-19-2-1004. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein. ",
+        "bbox": [
+            174,
+            789,
+            825,
+            872
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References   \n[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 3   \n[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. ArXiv preprint, abs/2204.14198, 2022. 3, 7, 8   \n[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3   \n[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 2425–2433. IEEE Computer Society, 2015. 7, 8   \n[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan, 2005. Association for Computational Linguistics. 6   \n[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 1, 3, 5   \n[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 190–200, Portland, Oregon, USA, 2011. Association for Computational Linguistics. 6   \n[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104–120. Springer, 2020. 2   \n[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 1931–1942. PMLR, 2021. 3   \n[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248–255. IEEE Computer Society, 2009. 4   \n[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. ArXiv preprint, abs/2106.11097, 2021. 3   \n[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6201–6210. IEEE, 2019. 3   \n[13] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot learners. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3816–3830, Online, 2021. Association for Computational Linguistics. 2   \n[14] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. AllenNLP: A deep semantic natural language processing platform. In Proceedings of Workshop for NLP Open Source Software (NLP-OSS), pages 1–6, Melbourne, Australia, 2018. Association for Computational Linguistics. 2 ",
+        "bbox": [
+            171,
+            65,
+            828,
+            914
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "[15] Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh, Lucy Vanderwende, Michel Galley, and Margaret Mitchell. Visual storytelling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1233–1239, San Diego, California, 2016. Association for Computational Linguistics. 2 ",
+        "bbox": [
+            173,
+            92,
+            826,
+            167
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[16] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976–12985, 2021. 2 ",
+        "bbox": [
+            171,
+            178,
+            820,
+            217
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 4904–4916. PMLR, 2021. 2 ",
+        "bbox": [
+            174,
+            224,
+            825,
+            289
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 5583–5594. PMLR, 2021. 2 ",
+        "bbox": [
+            173,
+            297,
+            825,
+            349
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32–73, 2017. 4 ",
+        "bbox": [
+            173,
+            358,
+            823,
+            409
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956–1981, 2020. 4 ",
+        "bbox": [
+            171,
+            417,
+            823,
+            458
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learningvia sparse sampling. In CVPR, 2021. 1, 3 ",
+        "bbox": [
+            171,
+            465,
+            823,
+            492
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[22] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. TVQA: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1369–1379, Brussels, Belgium, 2018. Association for Computational Linguistics. 3 ",
+        "bbox": [
+            173,
+            501,
+            823,
+            540
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[23] Jie Lei, Licheng Yu, Tamara Berg, and Mohit Bansal. What is more likely to happen next? videoand-language future event prediction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8769–8784, Online, 2020. Association for Computational Linguistics. 1, 6, 8 ",
+        "bbox": [
+            174,
+            549,
+            825,
+            599
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[24] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online, 2020. Association for Computational Linguistics. 1 ",
+        "bbox": [
+            173,
+            608,
+            825,
+            672
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[25] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In arxiv, 2021. 1, 3, 8, 9 ",
+        "bbox": [
+            171,
+            681,
+            823,
+            708
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ArXiv preprint, abs/2201.12086, 2022. 2, 3, 6, 7, 8 ",
+        "bbox": [
+            173,
+            717,
+            823,
+            756
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34, 2021. 2 ",
+        "bbox": [
+            173,
+            763,
+            823,
+            804
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[28] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: Hierarchical encoder for Video+Language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046–2065, Online, 2020. Association for Computational Linguistics. 1, 3 ",
+        "bbox": [
+            173,
+            811,
+            823,
+            863
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[29] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121–137. Springer, 2020. 2 ",
+        "bbox": [
+            174,
+            872,
+            826,
+            911
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[30] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL-04), pages 605–612, Barcelona, Spain, 2004. 6 ",
+        "bbox": [
+            171,
+            92,
+            823,
+            132
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6 ",
+        "bbox": [
+            173,
+            140,
+            821,
+            180
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[32] Xudong Lin, Gedas Bertasius, Jue Wang, Shih-Fu Chang, Devi Parikh, and Lorenzo Torresani. Vx2text: End-to-end learning of video-based text generation from multimodal inputs. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7001–7011. IEEE, 2021. 3, 4 ",
+        "bbox": [
+            173,
+            189,
+            823,
+            228
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853–13863, 2022. 3 ",
+        "bbox": [
+            171,
+            238,
+            825,
+            276
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[34] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang. Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. arXiv preprint arXiv:2206.02082, 2022. 1 ",
+        "bbox": [
+            173,
+            286,
+            825,
+            324
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[35] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13–23, 2019. 2 ",
+        "bbox": [
+            173,
+            334,
+            825,
+            398
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. ArXiv preprint, abs/2002.06353, 2020. 1, 3, 6 ",
+        "bbox": [
+            173,
+            409,
+            823,
+            446
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. ArXiv preprint, abs/2104.08860, 2021. 3 ",
+        "bbox": [
+            169,
+            457,
+            825,
+            483
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[38] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9876–9886. IEEE, 2020. 3 ",
+        "bbox": [
+            174,
+            492,
+            826,
+            544
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 2630–2640. IEEE, 2019. 2, 3 ",
+        "bbox": [
+            173,
+            554,
+            826,
+            606
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. ArXiv preprint, abs/2203.02155, 2022. 2, 6 ",
+        "bbox": [
+            171,
+            614,
+            823,
+            654
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 6 ",
+        "bbox": [
+            174,
+            662,
+            826,
+            715
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[42] Mandela Patrick, Po-Yao Huang, Yuki Markus Asano, Florian Metze, Alexander G. Hauptmann, João F. Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3 ",
+        "bbox": [
+            173,
+            724,
+            826,
+            776
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[43] Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao. Fewshot natural language generation for task-oriented dialog. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 172–182, Online, 2020. Association for Computational Linguistics. 1 [44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748–8763. PMLR, 2021. 2, 3, 4 ",
+        "bbox": [
+            171,
+            786,
+            826,
+            825
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            834,
+            826,
+            911
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. ArXiv preprint, abs/1910.10683, 2019. 1 ",
+        "bbox": [
+            171,
+            92,
+            823,
+            131
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[46] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1–67, 2020. 3 ",
+        "bbox": [
+            171,
+            141,
+            823,
+            180
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[47] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992, Hong Kong, China, 2019. Association for Computational Linguistics. 4, 6, 8 ",
+        "bbox": [
+            173,
+            190,
+            825,
+            242
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[48] Timo Schick and Hinrich Schütze. Exploiting cloze-questions for few-shot text classification and natural language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 255–269, Online, 2021. Association for Computational Linguistics. 2 ",
+        "bbox": [
+            173,
+            252,
+            823,
+            304
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[49] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. ArXiv preprint, abs/2201.08264, 2022. 1 ",
+        "bbox": [
+            171,
+            314,
+            825,
+            342
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[50] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: Plugging visual controls in text generation. ArXiv preprint, abs/2205.02655, 2022. 3 ",
+        "bbox": [
+            173,
+            352,
+            823,
+            391
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[51] Derek Tam, Rakesh R. Menon, Mohit Bansal, Shashank Srivastava, and Colin Raffel. Improving and simplifying pattern exploiting training. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4980–4991, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2 ",
+        "bbox": [
+            173,
+            401,
+            826,
+            453
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[52] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100–5111, Hong Kong, China, 2019. Association for Computational Linguistics. 2 ",
+        "bbox": [
+            174,
+            463,
+            826,
+            515
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[53] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200–212, 2021. 3 ",
+        "bbox": [
+            173,
+            525,
+            823,
+            564
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[54] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566–4575. IEEE Computer Society, 2015. 6 ",
+        "bbox": [
+            173,
+            575,
+            823,
+            614
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[55] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-tosequence learning framework. ArXiv preprint, abs/2202.03052, 2022. 3 ",
+        "bbox": [
+            171,
+            625,
+            823,
+            664
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[56] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation learning for text-video retrieval. ArXiv preprint, abs/2203.07111, 2022. 9 ",
+        "bbox": [
+            171,
+            674,
+            825,
+            702
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[57] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A largescale, high-quality multilingual dataset for video-and-language research. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 4580–4590. IEEE, 2019. 1, 6 ",
+        "bbox": [
+            173,
+            710,
+            826,
+            762
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[58] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. ArXiv preprint, abs/2108.10904, 2021. 2 ",
+        "bbox": [
+            169,
+            773,
+            823,
+            800
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. ArXiv preprint, abs/2109.01652, 2021. 5 ",
+        "bbox": [
+            171,
+            810,
+            825,
+            849
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[60] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017, pages 1645–1653, 2017. 1 ",
+        "bbox": [
+            174,
+            859,
+            825,
+            911
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[61] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1 [62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288–5296. IEEE Computer Society, 2016. 1, 6 [63] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. ArXiv preprint, abs/2109.05014, 2021. 1, 3, 4 [64] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian, Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework. ArXiv preprint, abs/2205.01818, 2022. 1 [65] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3208–3216, 2021. 2 [66] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. ArXiv preprint, abs/2111.11432, 2021. 2 [67] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. ArXiv preprint, abs/2201.02639, 2022. 1 [68] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 8 [69] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. ArXiv preprint, abs/2204.00598, 2022. 3, 4 [70] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579–5588, 2021. 2 [71] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. ArXiv preprint, abs/2205.01068, 2022. 1 [72] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In Marina Meila and Tong Zhang, editors, Proceedings of the   \n38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume   \n139 of Proceedings of Machine Learning Research, pages 12697–12706. PMLR, 2021. 1, 6 [73] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pre-training for image captioning and VQA. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 13041–13049. AAAI Press, 2020. 2 [74] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In Sheila A. McIlraith and Kilian Q. Weinberger, editors, Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 7590–7598. AAAI Press, 2018. 3, 6 [75] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. ArXiv preprint, abs/2112.01522, 2021. 3 ",
+        "bbox": [
+            171,
+            35,
+            828,
+            880
+        ],
+        "page_idx": 14
+    }
+]

parse/dev/shpkpVXzo3h/shpkpVXzo3h.md

@@ -0,0 +1,396 @@
+# 8-BIT OPTIMIZERS VIA BLOCK-WISE QUANTIZATION
+
+Anonymous authors Paper under double-blind review
+
+# ABSTRACT
+
+Stateful optimizers maintain gradient statistics over time, e.g., the exponentially smoothed sum (SGD with momentum) or squared sum (Adam) of past gradient values. This state can be used to accelerate optimization compared to plain stochastic gradient descent but uses memory that might otherwise be allocated to model parameters, thereby limiting the maximum size of models trained in practice. In this paper, we develop the first optimizers that use 8-bit statistics while maintaining the performance levels of using 32-bit optimizer states. To overcome the resulting computational, quantization, and stability challenges, we develop block-wise dynamic quantization. Block-wise quantization divides input tensors into smaller blocks that are independently quantized. Each block is processed in parallel across cores, yielding faster optimization and high precision quantization. To maintain stability and performance, we combine block-wise quantization with two additional changes: (1) dynamic quantization, a form of non-linear optimization that is precise for both large and small magnitude values, and (2) a stable embedding layer to reduce gradient variance that comes from the highly non-uniform distribution of input tokens in language models. As a result, our 8-bit optimizers maintain 32-bit performance with a small fraction of the memory footprint on a range of tasks, including 1.5B parameter language modeling, GLUE finetuning, ImageNet classification, WMT’14 machine translation, MoCo v2 contrastive ImageNet pretraining+finetuning, and RoBERTa pretraining, without changes to the original optimizer hyperparameters. We open-sourceour 8-bit optimizers as a drop-in replacement that only requires a two-line code change.
+
+Increasing model size is an effective way to achieve better performance for given resources (Kaplan et al., 2020; Henighan et al., 2020; Raffel et al., 2019; Lewis et al., 2021). However, training such large models requires storing the model, gradient, and state of the optimizer (e.g., exponentially smoothed sum and squared sum of previous gradients for Adam), all in a fixed amount of available memory. Although significant research has focused on enabling larger model training by reducing or efficiently distributing the memory required for the model parameters (Shoeybi et al., 2019; Lepikhin et al., 2020; Fedus et al., 2021; Brown et al., 2020; Rajbhandari et al., 2020), reducing the memory footprint of optimizer gradient statistics is much less studied. This is a significant missed opportunity since these optimizer states use $3 3 \mathrm { - } 7 5 \%$ of the total memory footprint during training. For example, the Adam optimizer states for the largest GPT-2 (Radford et al., 2019) and T5 (Raffel et al., 2019) models are 11 GB and 41 GB in size. In this paper, we develop a fast, high-precision non-linear quantization method – block-wise dynamic quantization – that enables stable 8-bit optimizers (e.g., Adam, AdamW, and Momentum) which maintain 32-bit performance at a fraction of the memory footprint and without any changes to the original hyperparameters.1
+
+While most current work uses 32-bit optimizer states, recent high-profile efforts to use 16-bit optimizers report difficultly for large models with more than 1B parameters (Ramesh et al., 2021). Going from 16-bit optimizers to 8-bit optimizers reduces the range of possible values from $2 ^ { 1 6 } = 6 5 5 3 \bar { 6 }$ values to just $\bar { 2 } ^ { 8 } = 2 5 6$ . To our knowledge, this has not been attempted before.
+
+Effectively using this very limited range is challenging for three reasons: quantization accuracy, computational efficiency, and large-scale stability. To maintain accuracy, it is critical to introduce some form of non-linear quantization to reduce errors for both common small magnitude values and rare large ones. However, to be practical, 8-bit optimizers need to be fast enough to not slow down training, which is especially difficult for non-linear methods that require more complex data structures to maintain the quantization buckets. Finally, to maintain stability with huge models beyond 1B parameters, a quantization method needs to not only have a good mean error but excellent worse case performance since a single large quantization error can cause the entire training run to diverge.
+
+![](images/48cb7db647012e734edb0dcae4a71dbc32cb6b6baab4efe52baf6bfa8d324e0c.jpg)  
+Figure 1: Schematic of 8-bit optimizers via block-wise dynamic quantization, see Section 2 for more details. After the optimizer update is performed in 32-bit, the state tensor is chunked into blocks, normalized by the absolute maximum value of each block. Then dynamic quantization is performed, and the index is stored. For dequantization, a lookup in the index is performed, with subsequent denormalization by multiplication with the block-wise absolute maximum value. Outliers are confined to a single block through block-wise quantization, and their effect on normalization is limited.
+
+We introduce a new block-wise quantization approach that addresses all three of these challenges. Block-wise quantization splits input tensors into blocks and performs quantization on each block independently. This block-wise division reduces the effect of outliers on the quantization process since they are isolated to particular blocks, thereby improving stability and performance, especially for large-scale models. Block-wise processing also allows for high optimizer throughput since each normalization can be computed independently in each core. This contrasts with tensor-wide normalization, which requires slow cross-core synchronization that is highly dependent on task-core scheduling. We combine block-wise quantization with two novel methods for stable, high-performance 8-bit optimizers: dynamic quantization and a stable embedding layer. Dynamic quantization is an extension of dynamic tree quantization for unsigned input data. The stable embedding layer is a variation of a standard word embedding layer that supports more aggressive quantization by normalizing the highly non-uniform distribution of inputs to avoid extreme gradient variation.
+
+Our 8-bit optimizers maintain 32-bit performance at a fraction of the original memory footprint. We show this for a broad range of tasks: 1.5B and $3 5 5 \mathrm { M }$ parameter language modeling, GLUE finetuning, ImageNet classification, WMT’ $^ { 1 4 + }$ WMT’16 machine translation, MoCo v2 contrastive image pretraining $^ +$ finetuning, and RoBERTa pretraining. We also report additional ablations and sensitivity analysis showing that all components – block-wise quantization, dynamic quantization, and stable embedding layer – are crucial for these results and that 8-bit Adam can be used as a simple drop-in replacement for 32-bit Adam, with no hyperparameter changes. We open-source our custom CUDA kernels and provide a PyTorch implementation that enables 8-bit optimization by changing two lines of code.
+
+# 1 BACKGROUND
+
+# 1.1 STATEFUL OPTIMIZERS
+
+An optimizer updates the parameters w of a neural network by using the gradient of the loss with respect to the weight $\begin{array} { r } { \mathbf { g } _ { t } \doteq \frac { \partial \mathbf { L } } { \partial \mathbf { w } } } \end{array}$ at update iteration $t$ . Stateful optimizers compute statistics of the gradient with respect to each parameter over time for accelerated optimization. Two of the most commonly used stateful optimizers are Adam (Kingma and Ba, 2014), and SGD with momentum (Qian, 1999) – or Momentum for short. Without damping and scaling constants, the update rules of these optimizers are given by:
+
+$$
+\begin{array} { r } { \mathbf { M o m e n t u m } ( \mathbf { g } _ { t } , \mathbf { w } _ { t - 1 } , \mathbf { m } _ { t - 1 } ) = \left\{ \begin{array} { l l } { \mathbf { m } _ { 0 } = \mathbf { g } _ { 0 } } \\ { \mathbf { m } _ { t } = \beta _ { 1 } \mathbf { m } _ { t - 1 } + \mathbf { g } _ { t } } \\ { \mathbf { w } _ { t } = \mathbf { w } _ { t - 1 } - { \boldsymbol \alpha } \cdot \mathbf { m } _ { t } } \end{array} \right. } \end{array}
+$$
+
+$$
+\mathrm { A d a m } ( \mathbf { g } _ { t } , \mathbf { w } _ { t - 1 } , \mathbf { m } _ { t - 1 } , \mathbf { r } _ { t - 1 } ) = \left\{ \begin{array} { l l } { \mathbf { r } _ { 0 } = \mathbf { m } _ { 0 } = \mathbf { 0 } } \\ { \mathbf { m } _ { t } = \beta _ { 1 } \mathbf { m } _ { t - 1 } + ( 1 - \beta _ { 1 } ) \mathbf { g } _ { t } } \\ { \mathbf { r } _ { t } = \beta _ { 2 } \mathbf { r } _ { t - 1 } + ( 1 - \beta _ { 2 } ) \mathbf { g } _ { t } ^ { 2 } } \\ { \mathbf { w } _ { t } = \mathbf { w } _ { t - 1 } - { \boldsymbol { \alpha } } \cdot \frac { \mathbf { m } _ { t } } { \sqrt { \mathbf { r } _ { t } } + \epsilon } } \end{array} \right.
+$$
+
+where $\beta _ { 1 }$ and $\beta _ { 2 }$ are smoothing constants, $\epsilon$ is a small constant, and $\alpha$ is the learning rate.
+
+For 32-bit states, Momentum and Adam consume 4 and 8 bytes per parameter. That is $_ 4 \mathrm { G B }$ and 8 GB for a 1B parameter model. Our 8-bit non-linear quantization reduces these costs to 1 GB and 2 GB.
+
+# 1.2 NON-LINEAR QUANTIZATION
+
+Quantization compresses numeric representations to save space at the cost of precision. Quantization is the mapping of a $k$ -bit integer to a real element in $D$ , that is, $\mathbf { Q } ^ { \mathrm { m a p } } \colon [ 0 , 2 ^ { k } - 1 ] \stackrel { \cdot } { \mapsto } D$ . For example, the IEEE 32-bit floating point data type maps the indices $0 . . . 2 ^ { 3 2 } - 1$ to the domain $\left[ - 3 . 4 \mathrm { e } 3 8 , + 3 . 4 \mathrm { e } 3 8 \right]$ . We use the following notation: $\mathbf { Q } ^ { \mathrm { m a p } } ( i ) = \mathbf { Q } _ { i } ^ { \mathrm { m a p } } = q _ { i }$ , for example $\mathbf { Q } ^ { \mathrm { m a p } } ( 2 ^ { 3 1 } + 1 3 1 0 7 2 ) = 2 . 0 3 1 2 5$ , for the IEEE 32-bit floating point data type.
+
+To perform general quantization from one data type into another we require three steps. (1) Compute a normalization constant $N$ that transforms the input tensor $\mathbf { T }$ into the range of the domain $D$ of the target quantization data type $\mathbf { Q } ^ { \mathrm { m a p } }$ , (2) for each element of $\mathbf { T } / N$ find the closest corresponding value $q _ { i }$ in the domain $D$ , (3) store the index $i$ corresponding to $q _ { i }$ in the quantized output tensor $\bar { \mathbf { T } } ^ { Q }$ . To receive the dequantized tensor $\mathbf { T } ^ { D }$ we look up the index and denormalize: $\mathbf { T } _ { i } ^ { D } = \mathbf { Q } ^ { \operatorname* { m a p } } ( \mathbf { T } _ { i } ^ { Q } ) \cdot N$ .
+
+To perform this procedure for dynamic quantization we first normalize into the range [-1, 1] through division by the absolute maximum value: $N = \operatorname* { m a x } ( | \mathbf { T } | )$ .
+
+Then we find the closest values via a binary search:
+
+$$
+\mathbf { T } _ { i } ^ { Q } = \underset { j = 0 } { \operatorname { a r g m i n } } | \mathbf { Q } _ { j } ^ { \operatorname* { m a p } } - \frac { \mathbf { T } _ { i } } { N } |
+$$
+
+# 1.3 DYNAMIC TREE QUANTIZATION
+
+Dynamic Tree quantization (Dettmers, 2016) is a method that yields low quantization error for both small and large magnitude values. Unlike data types with fixed exponent and fraction, dynamic tree quantization uses a datatype with a dynamic exponent and fraction that can change with each number. It is made up of four parts, as seen in Figure 2: (1) The first bit of the data type is reserved for a sign. (2) The number of subsequent zero bits indicates the magnitude of the exponent. (3) The first bit that is set to one indicates that all following values are reserved for (4) linear quantization. By moving the indicator bit, numbers can have a large exponent $1 0 ^ { - 7 }$ or precision as high as $1 / 6 3$ . Compared to linear quantization, dynamic tree quantization has better absolute and relative quantization errors for non-uniform distributions. Dynamic tree quantization is strictly defined to quantize numbers in the range [-1.0, 1.0], which is ensured by performing tensor-level absolute max normalization.
+
+![](images/1addf759c0e839af608df6324e894dbd84d099a00efed1da1b91e3383e234c2d.jpg)  
+Figure 2: Dynamic tree quantization.
+
+# 2 8-BIT OPTIMIZERS
+
+Our 8-bit optimizers have three components: (1) block-wise quantization that isolates outliers and distributes the error more equally over all bits; (2) dynamic quantization, which quantizes both small and large values with high precision; and (3) a stable embedding layer to improve stability during optimization for models with word embeddings.
+
+With these components, performing an optimizer update with 8-bit states is straightforward. We dequantize the 8-bit optimizer states to 32-bit, perform the update, and then quantize the states back to 8-bit for storage. We do this 8-bit to 32-bit conversion element-by-element in registers, which means no slow copies to GPU memory or additional temporary memory are needed to perform quantization and dequantization. For GPUs, this makes 8-bit optimizers faster than regular 32-bit optimizers, as we show in Section 3.
+
+# 2.1 BLOCK-WISE QUANTIZATION
+
+Our block-wise quantization reduces the cost of computing normalization and improves quantization precision by isolating outliers. In order to dynamically quantize a tensor, as defined in Section 1.2, we need to normalize the tensor into the range [-1, 1]. Such normalization requires a reduction over the entire tensor, which entails multiple synchronizations across GPU cores. Block-wise dynamic quantization reduces this cost by chunking an input tensor into small blocks of size $B = 2 0 4 8$ and performing normalization independently in each core across this block.
+
+More formally, using the notation introduced in Section 1.2, in block-wise quantization, we treat $\mathbf { T }$ as a one-dimensional sequence of elements that we chunk in blocks of size $B$ . This means for an input tensor $\mathbf { T }$ with $n$ elements we have $n / B$ blocks. We proceed to compute a normalization constant for each block: $N _ { b } = \operatorname* { m a x } ( | \mathbf { T } _ { b } | )$ , where $b$ is the index of the block $0 . . n / B$ . With this block-wise normalization constant, each block can be quantized independently:
+
+$$
+\mathbf { T } _ { b i } ^ { Q } = \underset { j = 0 } { \arg \operatorname* { m i n } } | \mathbf { Q } _ { j } ^ { \operatorname* { m a p } } - \frac { \mathbf { T } _ { b i } } { N _ { b } } | \bigg | _ { 0 < i < B }
+$$
+
+This approach has several advantages, both for stability and efficiency. First, each block normalization can be computed independently. Thus no synchronization between cores is required, and throughput is enhanced.
+
+Secondly, it is also much more robust to outliers in the input tensor. For example, to contrast blockwise and regular quantization, if we create an input tensor with one million elements sampled from the standard normal distribution, we expect less than $1 \%$ of elements of the tensor will be in the range $[ 3 , + \infty )$ . However, since we normalize the input tensor into the range [-1,1] this means the maximum values of the distribution determine the range of quantization buckets. This means if the input tensor contains an outlier with magnitude 5, the quantization buckets reserved for numbers between 3 and 5 will mostly go unused since less than $1 \%$ of numbers are in this range. With blockwise quantization, the effect of outliers is limited to a single block. As such, most bits are used effectively in other blocks.
+
+Furthermore, because outliers represent the absolute maximum value in the input tensor, blockwise quantization approximates outlier values without any error. This guarantees that the largest optimizer states, arguably the most important, will always be quantized with full precision. This property makes block-wise dynamic quantization both robust and precise and is essential for good training performance in practice.
+
+# 2.2 DYNAMIC QUANTIZATION
+
+In this work, we extend dynamic tree quantization (Section 1.3) for non-signed input tensors by re-purposing the sign bit. Since the second Adam state is strictly positive, the sign bit is not needed. Instead of just removing the sign bit, we opt to extend dynamic tree quantization with a fixed bit for the fraction. This extension is motivated by the observation that the second Adam state varies around 3-5 orders of magnitude during the training of a language model. In comparison, dynamic tree quantization already has a range of 7 orders of magnitude. We refer to this quantization as dynamic quantization to distinguish it from dynamic tree quantization in our experiments. A study of additional quantization data types and their performance is detailed in Appendix E.
+
+# 2.3 STABLE EMBEDDING LAYER
+
+Our stable embedding layer is a standard word embedding layer variation (Devlin et al., 2019) designed to ensure stable training for NLP tasks. This embedding layer supports more aggressive quantization by normalizing the highly non-uniform distribution of inputs to avoid extreme gradient variation. See Appendix B for a discussion of why commonly adopted embedding layers (Ott et al., 2019) are so unstable.
+
+We initialize the Stable Embedding layer with Xavier uniform initialization (Glorot and Bengio, 2010) and apply layer normalization (Ba et al., 2016) before adding position embeddings. This method maintains a variance of roughly one both at initialization and during training. Additionally, the uniform distribution initialization has less extreme values than a normal distribution, reducing maximum gradient size. Like Ramesh et al. (2021), we find that the stability of training improves significantly if we use 32-bit optimizer states for the embedding layers. This is the only layer that uses 32-bit optimizer states. We still use the standard precision for weights and gradients for the embedding layers – usually 16-bit. We show in our Ablation Analysis in Section 4 that this change is a necessary detail.
+
+# 3 8-BIT VS 32-BIT OPTIMIZER PERFORMANCE FOR COMMON BENCHMARKS
+
+Experimental Setup We compare the performance of 8-bit optimizers to their 32-bit counterparts on a range of challenging public benchmarks. These benchmarks either use Adam (Kingma and Ba, 2014), AdamW (Loshchilov and Hutter, 2018), or Momentum (Qian, 1999).
+
+We do not change any hyperparameters or precision of weights, gradients, and activations/input gradients for each experimental setting compared to the public baseline– the only change is to replace 32-bit optimizers with 8-bit optimizers. This means that for most experiments, we train in 16-bit mixed-precision (Micikevicius et al., 2017). We also compare with Adafactor (Shazeer and Stern, 2018), with the time-independent formulation for $\beta _ { 2 }$ (Shazeer and Stern, 2018) – which is the same formulation used in Adam. We also do not change any hyperparameters for Adafactor.
+
+We report on benchmarks in neural machine translation (Ott et al., 2018)2 trained on WMT’16 (Sennrich et al., 2016) and evaluated on en-de WMT’14 (Macha´cek and Bojar, 2014), large-scale ˇ language modeling (Lewis et al., 2021; Brown et al., 2020) and RoBERTa pretraining (Liu et al., 2019) on English CC- $1 0 0 +$ RoBERTa corpus (Nagel, 2016; Gokaslan and Cohen, 2019; Zhu et al., 2015; Wenzek et al., 2020), finetuning the pretrained masked language model RoBERTa (Liu et al., 2019)3 on GLUE (Wang et al., 2018a), ResNet-50 v1.5 image classification (He et al., 2016)4 on ImageNet-1k (Deng et al., 2009), and Moco v2 contrastive image pretraining and linear finetuning (Chen et al., $2 0 2 0 \mathrm { { b } } ) ^ { 5 }$ on ImageNet-1k (Deng et al., 2009).
+
+We use the stable embedding layer for all NLP tasks except for finetuning on GLUE. Beyond this, we follow the exact experimental setup outlined in the referenced papers and codebases. We consistently report replication results for each benchmark with public codebases and report median accuracy, perplexity, or BLEU over ten random seeds for GLUE, three random seeds for others tasks, and a single random seed for large scale language modeling. While it is standard to report means and standard errors on some tasks, others use median performance. We opted to report medians for all tasks for consistency.
+
+Results In Table 1, we see that 8-bit optimizers match replicated 32-bit performance for all tasks. While Adafactor is competitive with 8-bit Adam, 8-bit Adam uses less memory and provides faster optimization. Our 8-bit optimizers save up to $8 . 5 \mathrm { \ G B }$ of GPU memory for our largest 1.5B parameter language model and 2.0 GB for RoBERTa. Thus, 8-bit optimizers maintain performance and improve accessibility to the finetuning of large models for those that cannot afford GPUs with large memory buffers. We show models that are now accessible with smaller GPUs in Table 2. A breakdown of individual dataset results on GLUE can be found in Appendix A).
+
+Table 1: Median performance on diverse NLP and computer vision tasks: GLUE, object classification with (Moco v2) and without pretraining (CLS), machine translation (MT), and large-scale language modeling (LM). While 32-bit Adafactor is competitive with 8-bit Adam, it uses almost twice as much memory and trains slower. 8-bit Optimizers match or exceed replicated 32-bit performance on all tasks. We observe no instabilities for 8-bit optimizers. Time is total GPU time on V100 GPUs, except for RoBERTa and GPT3 pretraining, which were done on A100 GPUs.   
+
+<table><tr><td>Optimizer</td><td>Task</td><td>Data</td><td>Model</td><td>Metrict</td><td>Time</td><td>Mem saved</td></tr><tr><td>32-bit AdamW</td><td>GLUE</td><td>Multiple</td><td>RoBERTa-Large</td><td>88.9</td><td>1</td><td>Reference</td></tr><tr><td>32-bit AdamW</td><td>GLUE</td><td>Multiple</td><td>RoBERTa-Large</td><td>88.6</td><td>17h</td><td>0.0 GB</td></tr><tr><td>32-bit Adafactor</td><td>GLUE</td><td>Multiple</td><td>RoBERTa-Large</td><td>88.7</td><td>24h</td><td>1.3 GB</td></tr><tr><td>8-bit AdamW</td><td>GLUE</td><td>Multiple</td><td>RoBERTa-Large</td><td>88.7</td><td>15h</td><td>2.0 GB</td></tr><tr><td>32-bit Momentum</td><td>CLS</td><td>ImageNet-1k</td><td>ResNet-50</td><td>77.1</td><td>1</td><td>Reference</td></tr><tr><td>32-bit Momentum</td><td>CLS</td><td>ImageNet-1k</td><td>ResNet-50</td><td>77.1</td><td>118h</td><td>0.0 GB</td></tr><tr><td>8-bit Momentum</td><td>CLS</td><td>ImageNet-1k</td><td>ResNet-50</td><td>77.2</td><td>116 h</td><td>0.1 GB</td></tr><tr><td>32-bit Adam</td><td>MT</td><td>WMT&#x27;14+16</td><td>Transformer</td><td>29.3</td><td></td><td>Reference</td></tr><tr><td>32-bit Adam</td><td>MT</td><td>WMT&#x27;14+16</td><td>Transformer</td><td>29.0</td><td>126h</td><td>0.0 GB</td></tr><tr><td>32-bit Adafactor</td><td>MT</td><td>WMT&#x27;14+16</td><td>Transformer</td><td>29.0</td><td>127h</td><td>0.3 GB</td></tr><tr><td>8-bit Adam</td><td>MT</td><td>WMT&#x27;14+16</td><td>Transformer</td><td>29.1</td><td>115h</td><td>1.1 GB</td></tr><tr><td>32-bit Momentum</td><td>MoCo v2</td><td>ImageNet-1k</td><td>ResNet-50</td><td>67.5</td><td>1</td><td>Reference</td></tr><tr><td>32-bit Momentum</td><td>MoCo v2</td><td>ImageNet-1k</td><td>ResNet-50</td><td>67.3</td><td>30 days</td><td>0.0 GB</td></tr><tr><td>8-bit Momentum</td><td>MoCo v2</td><td>ImageNet-1k</td><td>ResNet-50</td><td>67.4</td><td>28 days</td><td>0.1 GB</td></tr><tr><td>32-bit Adam</td><td>LM</td><td>Multiple</td><td>Transformer-1.5B</td><td>9.0</td><td>308 days</td><td>0.0 GB</td></tr><tr><td>32-bit Adafactor</td><td>LM</td><td>Multiple</td><td>Transformer-1.5B</td><td>8.9</td><td>316days</td><td>5.6 GB</td></tr><tr><td>8-bit Adam</td><td>LM</td><td>Multiple</td><td>Transformer-1.5B</td><td>9.0</td><td>297 days</td><td>8.5 GB</td></tr><tr><td>32-bit Adam</td><td>LM</td><td>Multiple</td><td>GPT3-Medium</td><td>10.62</td><td>795 days</td><td>0.0 GB</td></tr><tr><td>32-bit Adafactor</td><td>LM</td><td>Multiple</td><td>GPT3-Medium</td><td>10.68</td><td>816 days</td><td>1.5 GB</td></tr><tr><td>8-bit Adam</td><td>LM</td><td>Multiple</td><td>GPT3-Medium</td><td>10.62</td><td>761 days</td><td>1.7 GB</td></tr><tr><td>32-bit Adam</td><td>Masked-LM</td><td>Multiple</td><td>RoBERTa-Base</td><td>3.49</td><td>101 days</td><td>0.0 GB</td></tr><tr><td>32-bit Adafactor</td><td>Masked-LM</td><td>Multiple</td><td>RoBERTa-Base</td><td>3.59</td><td>112 days</td><td>0.7 GB</td></tr><tr><td>8-bit Adam</td><td>Masked-LM</td><td>Multiple</td><td>RoBERTa-Base</td><td>3.48</td><td>94 days</td><td>1.1 GB</td></tr></table>
+
+†Metric: GLUE=Mean Accuracy/Correlation. CLS/MoCo $=$ Accuracy. MT $\sqsupseteq$ BLEU. LM=Perplexity.
+
+The broad range of tasks and competitive results demonstrate that 8-bit optimizers are a robust and effective replacement for 32-bit optimizers, do not require any additional changes in hyperparameters, and save a significant amount of memory while speeding up training slightly.
+
+Table 2: With 8-bit optimizers, larger models can be finetuned with the same GPU memory compared to standard 32-bit optimizer training. We use a batch size of one for this comparison.   
+
+<table><tr><td rowspan="2">GPU size in GB</td><td colspan="2">Largest finetunable Model (parameters)</td></tr><tr><td>32-bit Adam</td><td>8-bit Adam</td></tr><tr><td>6</td><td>RoBERTa-base (110M)</td><td>RoBERTa-large (355M)</td></tr><tr><td>11</td><td>MT5-small (300M)</td><td>MT5-base (580M)</td></tr><tr><td>24</td><td>MT5-base (580M)</td><td>MT5-large (1.2B)</td></tr><tr><td>24</td><td>GPT-2-medium (762M)</td><td>GPT-2-large (1.5B)</td></tr></table>
+
+# 4 ANALYSIS
+
+We analyze our method in two ways. First, we ablate all 8-bit optimizer components and show that they are necessary for good performance. Second, we look at the sensitivity to hyperparameters compared to 32-bit Adam and show that 8-bit Adam with block-wise dynamic quantization is a reliable replacement that does not require further hyperparameter tuning.
+
+Experimental Setup We perform our analysis on a strong 32-bit Adam baseline for language modeling with transformers (Vaswani et al., 2017). We subsample from the RoBERTa corpus (Liu et al., 2019) which consists of the English sub-datasets: Books (Zhu et al., 2015), Stories (Trinh and Le, 2018), OpenWebText-1 (Gokaslan and Cohen, 2019), Wikipedia, and CC-News (Nagel, 2016). We use a 50k token BPE encoded vocabulary (Sennrich et al., 2015). We find the best 2-GPU-day transformer baseline for 32-bit Adam with multiple hyperparameter searches that take in a total of 440 GPU days. Key hyperparameters include 10 layers with a model dimension of 1024, a fully connected hidden dimension of 8192, 16 heads, and input sub-sequences with a length of 512 tokens each. The final model has $2 0 9 \mathrm { m }$ parameters.
+
+Table 3: Ablation analysis of 8-bit Adam for small (2 GPU days) and large-scale ( ${ \approx } 1$ GPU year) transformer language models on the RoBERTa corpus. The runs without dynamic quantization use linear quantization. The percentage of unstable runs indicates either divergence or crashed training due to exploding gradients. We report median perplexity for successful runs. We can see that dynamic quantization is critical for general stability and block-wise quantization is critical for largescale stability. The stable embedding layer is useful for both 8-bit and 32-bit Adam and enhances stability to some degree.   
+
+<table><tr><td>Parameters</td><td>Optimizer</td><td>Dynamic</td><td>Block-wise</td><td>Stable Emb</td><td>Unstable (%)</td><td>Perplexity</td></tr><tr><td rowspan="8">209M</td><td>32-bit Adam</td><td></td><td></td><td></td><td>0</td><td>16.7</td></tr><tr><td>32-bit Adam</td><td></td><td></td><td>√</td><td>0</td><td>16.3</td></tr><tr><td>8-bit Adam</td><td></td><td></td><td></td><td>90</td><td>253.0</td></tr><tr><td>8-bit Adam</td><td></td><td></td><td>√</td><td>50</td><td>194.4</td></tr><tr><td>8-bit Adam</td><td></td><td></td><td></td><td>10</td><td>18.6</td></tr><tr><td>8-bit Adam</td><td>V</td><td></td><td>√</td><td>0</td><td>17.7</td></tr><tr><td>8-bit Adam</td><td>√</td><td>√</td><td></td><td>0</td><td>16.8</td></tr><tr><td>8-bit Adam</td><td>√</td><td>√</td><td>√</td><td>0</td><td>16.4</td></tr><tr><td>1.3B</td><td>32-bit Adam</td><td></td><td></td><td></td><td>0</td><td>10.4</td></tr><tr><td>1.3B</td><td>8-bit Adam</td><td>√</td><td></td><td></td><td>100</td><td>N/A</td></tr><tr><td>1.3B</td><td>8-bit Adam</td><td>√</td><td></td><td>√</td><td>80</td><td>10.9</td></tr><tr><td>1.5B</td><td>32-bit Adam</td><td></td><td></td><td></td><td>0</td><td>9.0</td></tr><tr><td>1.5B</td><td>8-bit Adam</td><td>√</td><td>一</td><td>√</td><td>0</td><td>9.0</td></tr></table>
+
+Ablation Analysis For the ablation analysis, we compare small and large-scale language modeling perplexity and training stability against a 32-bit Adam baseline. We ablate components individually and include combinations of methods that highlight their interactions. The baseline method uses linear quantization, and we add dynamic quantization, block-wise quantization, and the stable embedding layer to demonstrate their effect. To test optimization stability for small-scale language modeling, we run each setting with different hyperparameters and report median performance across all successful runs. A successful run is a run that does not crash due to exploding gradients or diverges in the loss. We use the hyperparameters $\epsilon$ {1e-8, 1e-7, 1e-6}, $\beta _ { 1 }$ $\{ 0 . 9 0 , \ 0 . 8 7 , \ 0 . 9 3 \}$ , $\beta _ { 2 }$ $\{ 0 . 9 9 9 , 0 . 9 9 , 0 . 9 8 \}$ and small changes in learning rates. We also include some partial ablations for large-scale models beyond 1B parameters. In the large-scale setting, we run several seeds with the same hyperparameters. We use a single seed for 32-bit Adam, five seeds for 8-bit Adam at 1.3B parameters, and a single seed for 8-bit Adam at 1.5B parameters.6 Results are shown in Table 3.
+
+The Ablations show that dynamic quantization, block-wise quantization, and the stable embedding layer are critical for either performance or stability. In addition, block-wise quantization is critical for large-scale language model stability.
+
+Sensitivity Analysis We compare the perplexity of 32-bit Adam vs 8-bit Adam $^ +$ Stable Embedding as we change the optimizer hyperparameters: learning rate, betas, and $\epsilon$ . We change each hyperparameter individually from the baseline hyperparameters $\beta _ { 1 } { = } 0 . 9$ , $\beta _ { 2 } { = } 0 . 9 9 5$ , $\scriptstyle \epsilon = 1 \mathrm { e } - 7$ , and $_ { \mathrm { l r = 0 . 0 1 6 3 } }$ and run two random seeds for both 8-bit and 32-bit Adam for each setting. If 8-bit Adam is perfectly insensitive to hyperparameters compared to 32-bit Adam, we would expect the same constant offset in performance for any hyperparameter combination. The results can be seen in Figure 3. The results show a relatively steady gap between 8-bit and 32-bit Adam, suggesting that 8-bit Adam does not require any further hyperparameter tuning compared to 32-bit Adam.
+
+![](images/a491a7b429b46d1a19aa6395663d42a67478d9b6d41c62e20cb6b0a648a555f0.jpg)  
+Figure 3: Sensitivity analysis of 8-bit vs 32-bit Adam hyperparameters. We can see that there is little variance between 8 and 32-bit performance, which suggests that 8-bit Adam can be used as a drop-in replacement for 32-bit Adam without any further hyperparameter tuning.
+
+# 5 RELATED WORK
+
+Compressing & Distributing Optimizer States While 16-bit Adam has been used in several publications, the stability of 16-bit Adam was first explicitly studied for a text-to-image generation model DALL-E (Ramesh et al., 2021). They show that a stable embedding layer, tensor-wise scaling constants for both Adam states, and multiple loss scaling blocks are critical to achieving stability during training. Our work reduces the memory footprint of Adam further, from 16 to 8-bit. In addition, we achieve stability by developing new training procedures and non-linear quantization, both of which complement previous developments.
+
+Adafactor (Shazeer and Stern, 2018) uses a different strategy to save memory. All optimizer states are still 32-bit, but the second Adam state is factorized by a row-column outer product resulting in a comparable memory footprint to 16-bit Adam. Alternatively, Adafactor can also be used without using the first moment $\beta _ { 1 } = 0 . 0$ ) (Lepikhin et al., 2020). This version is as memory efficient as 8-bit Adam, but unlike 8-bit Adam, hyperparameters for this Adafactor variant need to be re-tuned to achieve good performance. We compare 8-bit Adam with Adafactor $\beta _ { 1 } > 0 . 0$ in our experiments.
+
+AdaGrad (Duchi et al., 2011) adapts the gradient with aggregate training statistics over the entire training run. AdaGrad that uses only the main diagonal as optimizer state and extensions of AdaGrad such as SM3 (Anil et al., 2019) and extreme tensoring (Chen et al., 2020a) can be more efficient than 8-bit Adam. We include some initial comparison with AdaGrad in Appendix G.
+
+Optimizer sharding (Rajbhandari et al., 2020) splits optimizer states across multiple accelerators such as GPUs/TPUs. While very effective, it can only be used if multiple accelerators are available and data parallelism is used. Optimizer sharding can also have significant communication overhead (Rajbhandari et al., 2021). Our 8-bit optimizers work with all kinds of parallelism. They can also complement optimizer sharding, as they reduce communication overhead by $7 5 \%$ .
+
+General Memory Reduction Techniques Other complementary methods for efficient training can be either distributed or local. Distributed approaches spread out the memory of a model across several accelerators such as GPUs/TPUs. Such approaches are model parallelism (Krizhevsky et al., 2009), pipeline parallelism (Krizhevsky et al., 2009; Huang et al., 2018; Harlap et al., 2018), and operator parallelism (Lepikhin et al., 2020). These approaches are useful if one has multiple accelerators available. Our 8-bit optimizers are useful for both single and multiple devices.
+
+Local approaches work for a single accelerator. They include gradient checkpointing (Chen et al., 2016), reversible residual connections (Gomez et al., 2017), and offloading (Pudipeddi et al., 2020;
+
+Rajbhandari et al., 2021). All these methods save memory at the cost of increased computational or communication costs. Our 8-bit optimizers reduce the memory footprint of the model while maintaining 32-bit training speed.
+
+Quantization Methods and Data Types While our work is the first to apply 8-bit quantization to optimizer statistics, quantization for neural network model compression, training, and inference are well-studied problems. One of the most common formats of 8-bit quantization is to use data types composed of static sign, exponent, and fraction bits. The most common combination is 5 bits for the exponent and 2 bits for the fraction (Wang et al., 2018b; Sun et al., 2019; Cambier et al., 2020; Mellempudi et al., 2019) with either no normalization or min-max normalization. These data types offer high precision for small magnitude values but have large errors for large magnitude values since only 2 bits are assigned to the fraction. Other methods improve quantization through soft constraints (Li et al., 2021) or more general uniform affine quantizations (Pappalardo, 2021).
+
+Data types lower than 8-bit are usually used to prepare a model for deployment, and the main focus is on improving network inference speed and memory footprint rather than maintaining accuracy. There are methods that use 1-bit (Courbariaux and Bengio, 2016; Rastegari et al., 2016; Courbariaux et al., 2015), 2-bit/3 values (Zhu et al., 2017; Choi et al., 2019), 4-bits (Li et al., 2019), more bits (Courbariaux et al., 2014), or a variable amount of bits (Gong et al., 2019). See also Qin et al. (2020) for a survey on binary neural networks. While these low-bit quantization techniques allow for efficient storage, they likely lead to instability when used for optimizer states.
+
+The work most similar to our block-wise quantization is work on Hybrid Block Floating Point (HBFP) (Drumond et al., 2018) which uses a 24-bit fraction data type with a separate exponent for each tile in matrix multiplication to perform 24-bit matrix multiplication. However, unlike HBFP, block-wise dynamic quantization has the advantage of having both block-wise normalization and a dynamic exponent for each number. This allows for a much broader range of important values since optimizer state values vary by about 5 orders of magnitude. Furthermore, unlike HBFP, block-wise quantization approximates the maximum magnitude values within each block without any quantization error, which is critical for optimization stability, particularly for large networks.
+
+# 6 DISCUSSION & LIMITATIONS
+
+Here we have shown that high precision quantization can yield 8-bit optimizers that maintain 32-bit optimizer performance without requiring any change in hyperparameters. One of the main limitations of our work is that 8-bit optimizers for natural language tasks require a stable embedding layer to be trained to 32-bit performance. On the other hand, we show that 32-bit optimizers also benefit from a stable embedding layer. As such, the stable embedding layer could be seen as a general replacement for other embedding layers.
+
+We show that 8-bit optimizers reduce the memory footprint and accelerate optimization on a wide range of tasks. However, since 8-bit optimizers reduce only the memory footprint proportional to the number of parameters, models that use large amounts of activation memory and little memory for parameters, such as convolutional networks, have few benefits from using 8-bit optimizers. Thus, 8- bit optimizers are most beneficial for training or finetuning models with many parameters on highly memory-constrained GPUs.
+
+Furthermore, there remain sources of instability that, to our knowledge, are not well understood. For example, we observed that models with over 1B parameters often have hard systemic divergence, where many parameters simultaneously cause exploding gradients. In other cases, a single parameter among those 1B parameters assumed a value too large, caused an exploding gradient, and led to a cascade of instability. It might be that this rare, soft cascading instability is related to the phenomena where instability disappears after reloading a model checkpoint and rolling a new random seed – a method standard for training huge models. This cascading instability might also be related to the observation that the larger a model is, the more unstable it becomes. For our 8-bit optimizers, we primarily needed the stable embedding layer to avoid cascading instability. Thus the stable embedding layer could potentially be viewed as decreasing the probability of extreme outlier gradients. If such phenomena were better understood, it could lead to better 8-bit optimizers and more stable training in general.
+
+# REFERENCES
+
+Anil, R., Gupta, V., Koren, T., and Singer, Y. (2019). Memory efficient adaptive optimization. In Wallach, H. M., Larochelle, H., Beygelzimer, A., d’Alche-Buc, F., Fox, E. B., and Garnett, R., ´ editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 9746–9755.
+
+Ba, J. L., Kiros, J. R., and Hinton, G. E. (2016). Layer normalization. arXiv preprint arXiv:1607.06450.
+
+Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020). Language models are few-shot learners. arXiv preprint arXiv:2005.14165.
+
+Cambier, L., Bhiwandiwalla, A., Gong, T., Elibol, O. H., Nekuii, M., and Tang, H. (2020). Shifted and squeezed 8-bit floating point format for low-precision training of deep neural networks. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.
+
+Chen, E. J. and Kelton, W. D. (2001). Quantile and histogram estimation. In Proceeding of the 2001 Winter Simulation Conference (Cat. No. 01CH37304), volume 1, pages 451–459. IEEE.
+
+Chen, T., Xu, B., Zhang, C., and Guestrin, C. (2016). Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174.
+
+Chen, X., Agarwal, N., Hazan, E., Zhang, C., and Zhang, Y. (2020a). Extreme tensoring for lowmemory preconditioning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.
+
+Chen, X., Fan, H., Girshick, R., and He, K. (2020b). Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297.
+
+Choi, J., Venkataramani, S., Srinivasan, V., Gopalakrishnan, K., Wang, Z., and Chuang, P. (2019). Accurate and efficient 2-bit quantized neural networks. In Talwalkar, A., Smith, V., and Zaharia, M., editors, Proceedings of Machine Learning and Systems 2019, MLSys 2019, Stanford, CA, USA, March 31 - April 2, 2019. mlsys.org.
+
+Courbariaux, M. and Bengio, Y. (2016). Binarynet: Training deep neural networks with weights and activations constrained to $+ 1$ or -1. CoRR, abs/1602.02830.
+
+Courbariaux, M., Bengio, Y., and David, J. (2015). Binaryconnect: Training deep neural networks with binary weights during propagations. In Cortes, C., Lawrence, N. D., Lee, D. D., Sugiyama, M., and Garnett, R., editors, Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 3123–3131.
+
+Courbariaux, M., Bengio, Y., and David, J.-P. (2014). Training deep neural networks with low precision multiplications. arXiv preprint arXiv:1412.7024.
+
+Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee.
+
+Dettmers, T. (2016). 8-bit approximations for parallelism in deep learning. International Conference on Learning Representations (ICLR).
+
+Devlin, J., Chang, M., Lee, K., and Toutanova, K. (2019). BERT: pre-training of deep bidirectional transformers for language understanding. In Burstein, J., Doran, C., and Solorio, T., editors, Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics.
+
+Drumond, M., Lin, T., Jaggi, M., and Falsafi, B. (2018). Training dnns with hybrid block floating point. In Bengio, S., Wallach, H. M., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R., editors, Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montreal, ´ Canada, pages 451–461.
+
+Duchi, J., Hazan, E., and Singer, Y. (2011). Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7).
+
+Dunning, T. and Ertl, O. (2019). Computing extremely accurate quantiles using t-digests. arXiv preprint arXiv:1902.04023.
+
+Fedus, W., Zoph, B., and Shazeer, N. (2021). Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. arXiv preprint arXiv:2101.03961.
+
+Glorot, X. and Bengio, Y. (2010). Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings.
+
+Gokaslan, A. and Cohen, V. (2019). Openwebtext corpus.
+
+Gomez, A. N., Ren, M., Urtasun, R., and Grosse, R. B. (2017). The reversible residual network: Backpropagation without storing activations. arXiv preprint arXiv:1707.04585.
+
+Gong, R., Liu, X., Jiang, S., Li, T., Hu, P., Lin, J., Yu, F., and Yan, J. (2019). Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 4851–4860. IEEE.
+
+Govindaraju, N. K., Raghuvanshi, N., and Manocha, D. (2005). Fast and approximate stream mining of quantiles and frequencies using graphics processors. In Proceedings of the 2005 ACM SIGMOD international conference on Management of data, pages 611–622.
+
+Greenwald, M. and Khanna, S. (2001). Space-efficient online computation of quantile summaries. ACM SIGMOD Record, 30(2):58–66.
+
+Harlap, A., Narayanan, D., Phanishayee, A., Seshadri, V., Devanur, N., Ganger, G., and Gibbons, P. (2018). Pipedream: Fast and efficient pipeline parallel dnn training. arXiv preprint arXiv:1806.03377.
+
+He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778.
+
+Henighan, T., Kaplan, J., Katz, M., Chen, M., Hesse, C., Jackson, J., Jun, H., Brown, T. B., Dhariwal, P., Gray, S., et al. (2020). Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701.
+
+Huang, Y., Cheng, Y., Bapna, A., Firat, O., Chen, M. X., Chen, D., Lee, H., Ngiam, J., Le, Q. V., Wu, Y., et al. (2018). Gpipe: Efficient training of giant neural networks using pipeline parallelism. arXiv preprint arXiv:1811.06965.
+
+Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361.
+
+Keskar, N. S., McCann, B., Varshney, L. R., Xiong, C., and Socher, R. (2019). CTRL: A conditional transformer language model for controllable generation. CoRR, abs/1909.05858.
+
+Kingma, D. P. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
+
+Krizhevsky, A., Hinton, G., et al. (2009). Learning multiple layers of features from tiny images.
+
+Lepikhin, D., Lee, H., Xu, Y., Chen, D., Firat, O., Huang, Y., Krikun, M., Shazeer, N., and Chen, Z. (2020). Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668.
+
+Lewis, M., Bhosale, S., Dettmers, T., Goyal, N., and Zettlemoyer, L. (2021). Base layers: Simplifying training of large, sparse models. arXiv preprint arXiv:2103.16716.
+
+Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., and Zettlemoyer, L. (2020). BART: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Jurafsky, D., Chai, J., Schluter, N., and Tetreault, J. R., editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7871–7880. Association for Computational Linguistics.
+
+Li, J. B., Qu, S., Li, X., Strubell, E., and Metze, F. (2021). End-to-end quantized training via log-barrier extensions.
+
+Li, R., Wang, Y., Liang, F., Qin, H., Yan, J., and Fan, R. (2019). Fully quantized network for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2810–2819. Computer Vision Foundation / IEEE.
+
+Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V. (2019). Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
+
+Loshchilov, I. and Hutter, F. (2018). Fixing weight decay regularization in adam.
+
+Macha´cek, M. and Bojar, O. (2014). Results of the wmt14 metrics shared task. In ˇ Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 293–301.
+
+Mellempudi, N., Srinivasan, S., Das, D., and Kaul, B. (2019). Mixed precision training with 8-bit floating point. CoRR, abs/1905.12334.
+
+Micikevicius, P., Narang, S., Alben, J., Diamos, G., Elsen, E., Garcia, D., Ginsburg, B., Houston, M., Kuchaiev, O., Venkatesh, G., et al. (2017). Mixed precision training. arXiv preprint arXiv:1710.03740.
+
+Nagel, S. (2016). Cc-news.
+
+Ott, M., Edunov, S., Baevski, A., Fan, A., Gross, S., Ng, N., Grangier, D., and Auli, M. (2019). fairseq: A fast, extensible toolkit for sequence modeling. arXiv preprint arXiv:1904.01038.
+
+Ott, M., Edunov, S., Grangier, D., and Auli, M. (2018). Scaling neural machine translation. arXiv preprint arXiv:1806.00187.
+
+Pappalardo, A. (2021). Xilinx/brevitas.
+
+Pudipeddi, B., Mesmakhosroshahi, M., Xi, J., and Bharadwaj, S. (2020). Training large neural networks with constant memory using a new execution algorithm. arXiv preprint arXiv:2002.05645.
+
+Qian, N. (1999). On the momentum term in gradient descent learning algorithms. Neural networks : the official journal of the International Neural Network Society, 12 1:145–151.
+
+Qin, H., Gong, R., Liu, X., Bai, X., Song, J., and Sebe, N. (2020). Binary neural networks: A survey. CoRR, abs/2004.03333.
+
+Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.
+
+Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J. (2019). Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683.
+
+Rajbhandari, S., Rasley, J., Ruwase, O., and He, Y. (2020). Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–16. IEEE.   
+Rajbhandari, S., Ruwase, O., Rasley, J., Smith, S., and He, Y. (2021). Zero-infinity: Breaking the gpu memory wall for extreme scale deep learning. arXiv preprint arXiv:2104.07857.   
+Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., and Sutskever, I. (2021). Zero-shot text-to-image generation. arXiv preprint arXiv:2102.12092.   
+Rastegari, M., Ordonez, V., Redmon, J., and Farhadi, A. (2016). Xnor-net: Imagenet classification using binary convolutional neural networks. In Leibe, B., Matas, J., Sebe, N., and Welling, M., editors, Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV, volume 9908 of Lecture Notes in Computer Science, pages 525–542. Springer.   
+Sennrich, R., Haddow, B., and Birch, A. (2015). Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909.   
+Sennrich, R., Haddow, B., and Birch, A. (2016). Edinburgh neural machine translation systems for wmt 16. arXiv preprint arXiv:1606.02891.   
+Shazeer, N. and Stern, M. (2018). Adafactor: Adaptive learning rates with sublinear memory cost. In International Conference on Machine Learning, pages 4596–4604. PMLR.   
+Shoeybi, M., Patwary, M., Puri, R., LeGresley, P., Casper, J., and Catanzaro, B. (2019). Megatronlm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053.   
+Sun, X., Choi, J., Chen, C., Wang, N., Venkataramani, S., Srinivasan, V., Cui, X., Zhang, W., and Gopalakrishnan, K. (2019). Hybrid 8-bit floating point (HFP8) training and inference for deep neural networks. In Wallach, H. M., Larochelle, H., Beygelzimer, A., d’Alche-Buc, F., Fox, ´ E. B., and Garnett, R., editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 4901–4910.   
+Trinh, T. H. and Le, Q. V. (2018). A simple method for commonsense reasoning. arXiv preprint arXiv:1806.02847.   
+Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. (2017). Attention is all you need. arXiv preprint arXiv:1706.03762.   
+Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., and Bowman, S. R. (2018a). Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461.   
+Wang, N., Choi, J., Brand, D., Chen, C., and Gopalakrishnan, K. (2018b). Training deep neural networks with 8-bit floating point numbers. In Bengio, S., Wallach, H. M., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R., editors, Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montreal, Canada ´ , pages 7686–7695.   
+Wenzek, G., Lachaux, M.-A., Conneau, A., Chaudhary, V., Guzman, F., Joulin, A., and Grave, E. ´ (2020). CCNet: Extracting high quality monolingual datasets from web crawl data. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 4003–4012, Marseille, France. European Language Resources Association.   
+Zhu, C., Han, S., Mao, H., and Dally, W. J. (2017). Trained ternary quantization. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.   
+Zhu, Y., Kiros, R., Zemel, R., Salakhutdinov, R., Urtasun, R., Torralba, A., and Fidler, S. (2015). Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pages 19–27.
+
+# A GLUE SCORE BREAKDOWN
+
+Table 4 contains the breakdown of individual scores on the GLUE datasets.
+
+Table 4: Breakdown of GLUE scores. Each column is the median of 10 random seeds. The mean is the mean over medians.   
+
+<table><tr><td>Model</td><td>MNLI</td><td>QNLI</td><td>QQP</td><td>RTE</td><td>SST-2</td><td>MRPC</td><td>CoLA</td><td>STS-B</td><td>Mean</td></tr><tr><td>32-bit Adam</td><td>90.40</td><td>94.85</td><td>92.2</td><td>84.5</td><td>96.40</td><td>90.1</td><td>67.41</td><td>93.03</td><td>88.61</td></tr><tr><td>32-bit Adafactor</td><td>90.35</td><td>94.70</td><td>92.2</td><td>85.4</td><td>96.45</td><td>90.0</td><td>67.63</td><td>92.91</td><td>88.71</td></tr><tr><td>8-bit Adam</td><td>90.30</td><td>94.70</td><td>92.2</td><td>85.9</td><td>96.40</td><td>90.3</td><td>67.20</td><td>92.87</td><td>88.73</td></tr></table>
+
+# B STABILITY OF EMBEDDING LAYERS
+
+Highly variable gradients can lead to unpredictable optimization behavior and instability that manifests as divergence or exploding gradients. Low precision optimziers can amplify variance of gradient updates due to the noise introduced during quantization. While our 8-bit optimizers appear to be stable for convolutional networks, similar to Ramesh et al. (2021), we find that word embedding layers are a major source of instability.
+
+The main instability from the word embedding layer comes from the fact that it is a sparse layer with non-uniform distribution of inputs which can produce maximum gradient magnitudes $1 0 0 \mathrm { x }$ larger than other layers. For dense layers, if given $n$ samples arranged into $k$ mini-batches the sum of gradients of all mini-batches is always the same independent of how the $n$ samples are arranged into $k$ mini-batches. For embedding gradients, this depends on the arrangement of samples into mini-batches. This is because most deep learning frameworks normalize the gradient by the number of total tokens in the mini-batch, rather than the frequency of each individual token. This approximation allows stable learning with a single learning rate rather than variable learning rates that depend on token frequency in each individual mini-batch. However a side-effect of this method is that the magnitude of gradients for a particular token can vary widely with batch sizes and between different mini-batches.
+
+There are multiple recipes for initialization word embedding layers. One of the most common recipes used in all models trained with fairseq (Ott et al., 2019) such as RoBERTa (Liu et al., 2019), BART (Lewis et al., 2020), large NMT models (Ott et al., 2018), and sparse expert models (Lewis√ et al., 2021), is the following: Initialize the word embedding layer with $N ( 0 , 1 / { \sqrt { k } } )$ where $k$ is the embedding size of the embedding layer and to scale the outputs by $\sqrt { k }$ . This scheme has a variance of one at the start of training for the output distribution to ensure good gradient flow.
+
+We find this approach to induce some instability for 8-bit optimizers. We develop the stable embedding layer to solve this instability problem.
+
+While the full recipe for our stable embedding layer is new, components of it has been used before. The layer norm after the embedding has been used before in work such as Devlin et al. (2019) and Radford et al. (2019) and enhanced precision for this particular layer was used in Ramesh et al. (2021). As pointed out above, these elements are not standard and the stable embedding layer combines three aspects that are all important: (1) enhanced precision, (2) layer norm, and (3) Xavier initialization.
+
+# C QUANTIZATION ERROR ANALYSIS
+
+To gain more insights into why block-wise dynamic quantization works so well and how it could be improved, we performed a quantization error analysis of Adam quantization errors during language model training. Adam quantization errors are the deviations between the quantized 8-bit Adam update and the 32-bit Adam updates: $\lvert \mathbf { u 8 } - \mathbf { u _ { 1 6 } } \rvert$ , where $\mathbf { u } _ { \mathbf { k } } = \mathbf { s _ { 1 } ^ { k } } / \mathbf { s _ { 2 } ^ { k } }$ for $k$ bits. See Background Section 1.1 for details on Adam.
+
+A good 8-bit quantization has the property that, for a given input distribution, the inputs are only rarely quantized into intervals with high quantization error and most often quantized into intervals with low error.
+
+In 8-bit, there are $2 5 5 \times 2 5 6$ possible 8-bit Adam updates, 256 possible values for the first and 256 for the second Adam state. We look at the average quantization error of each of these possible updates to see where the largest errors are and we plot histograms to see how often do these values with high error occur. Taken together, these two perspectives give a detailed view of the magnitude of deviations and how often large deviations occur.
+
+We study these questions by looking at how often each of the 256 values for both Adam states are used during language model training. We also analyze the average error for each of the inputs quantized to each of the 256 values. With this analysis it is easy to find regions of high use and high error, and visualize their overlap. An overlap of these regions is associated with large frequent errors that cause unstable training. The quantization error analysis is shown in Figure 4.
+
+The plots show two things: (1) The region of high usage (histogram) shows how often each combination of $2 5 6 \times 2 5 6$ bit values is used for the first Adam state $\mathbf { s _ { 1 } }$ (exponentially smoothed running sum) and the second Adam state $\mathbf { s _ { 2 } }$ (exponentially smoothed running squared sum). (2) The error plots show for $k$ -bit Adam updates $\mathbf { u _ { k } } = \mathbf { s _ { 1 } } / ( \sqrt { \mathbf { s _ { 2 } } } + \epsilon )$ the mean absolute Adam error $| u _ { 3 2 } - u _ { 8 } |$ and the relative Adam error $| \boldsymbol { u } _ { 3 2 } - \boldsymbol { u } _ { 8 } | / | \boldsymbol { u } _ { 3 2 } |$ averaged over each bit combination. In conjunction these plots show which bits have the highest error per use and how often each bit is used. The $\mathbf { X }$ - axis/y-axis represents the quantization type range which means the largest positive/negative Adam states per block/tensor take the values $1 . 0 / - 1 . 0 $ .
+
+We can see that block-wise dynamic quantization has the smallest overlap between regions of high use and high error. While the absolute Adam quantization error of block-wise dynamic quantization is 0.0061, which is not much lower than that of dynamic quantization with 0.0067, the plots can also be interpreted as block-wise dynamic having rarer large errors that likely contribute to improved stability during optimization.
+
+# D FINE-GRAINED OPTIMIZER RUNTIME PERFORMANCE
+
+Table 5 shows optimizer performance that is benchmarked in isolation without any training. We use a large sample of a normal distribution and benchmark the average time to perform 100 optimizer updates per billion parameters in milliseconds.
+
+Table 5: Runtime performance of 8-bit optimizers vs commonly used 32-bit optimizers in milliseconds per update per 1B parameters for 32-bit gradients. This comparision was run on a V100 GPU.   
+
+<table><tr><td></td><td colspan="3">Milliseconds per update per 1B param</td></tr><tr><td>Optimizer</td><td>32-bit PyTorch</td><td>32-bit Apex</td><td>8-bit (Ours)</td></tr><tr><td>Adam</td><td>145</td><td>63</td><td>47</td></tr><tr><td>Momentum</td><td>58</td><td>46</td><td>34</td></tr><tr><td>LAMB</td><td></td><td>91</td><td>65</td></tr><tr><td>LARS</td><td>1</td><td>119</td><td>43</td></tr></table>
+
+# E ADDITIONAL QUANTIZATION DATA TYPES
+
+This section describes additional quantization data types that we tried but which we found to perform poorly in quantization performance or stability. While quantile quantization has an average quantization twice as low as dynamic quantization for any normal distribution it has sporadic large errors that lead to large Adam errors and poor model performance (see Figure 5) and even with state-of-the-art quantile estimation algorithms (see Section F) quantile quantization is too slow to be practical. An overview of quantization performance of this additional quantization data types compared to dynamic quantization (without block-wise quantization) can be found in Table 6.
+
+![](images/eb34d99c35d797a1f3ec0e6b74881b63bc240858d5694fea2e006b563e0edd2a.jpg)  
+Figure 4: Good quantization methods do not have overlaps between regions of high use and high error. The plot shows that for linear quantization regions of high usage and high error overlap. For dynamic quantization regions with high relative error are used infrequently while only small regions have high usage and high absolute error. Block-wise dynamic quantization spreads out the usage over a large space and has the lowest overlap between regions of high use and errors. This means that not only is the overall error of block-wise dynamic quantization lower, but also that large errors for individual parameter updates are rarer compared to other methods, thus improving stability. See the main text for more details.
+
+Table 6: Mean relative Adam and absolute quantization error for the first Adam state for different quantization methods. Results show mean±standard error. We can see that Dynamic Quantization has best relative error and that both Dynamic methods have the best absolute error.   
+
+<table><tr><td>Method</td><td>Relative Adam Error</td><td>Absolute Quantization Error</td></tr><tr><td>Linear</td><td>201% ±17%</td><td>41.2e-10±3.1e-10</td></tr><tr><td>Quantile</td><td>11.9% ± 0.3%</td><td>8.8e-10±0.9e-10</td></tr><tr><td>Inverse Dynamic</td><td>6.5%± 0.1%</td><td>4.6e-10±0.4e-10</td></tr><tr><td>Dynamic</td><td>4.8%± 0.4%</td><td>3.5e-10±1.1e-10</td></tr></table>
+
+![](images/1302c59d4d0e7dc6fd72b7500bf52e0c6114846c7c1b4c951a11a58a3bce8316.jpg)  
+Figure 5: Distribution of Adam error among each of the 256 8-bit values of the first Adam state. We normalize the values into the range [-1,1]. With this, -1 indicates the largest negative value, 0 the value that is closest to 0, and so forth. See Figure 6 for a visualization of this normalization. Quantile quantization has large errors for large values, while dynamic quantization has small errors for both small and large values while the bulk of the errors is concentrated in intermediate values.
+
+# E.1 INVERSE DYNAMIC QUANTIZATION
+
+Inverse Dynamic Quantization is motivated by the hypothesis that large Adam updates are more important than small updates. Since Adam is composed of a ratio of optimizer states $\mathbf { m } _ { t } / ( \sqrt { \mathbf { r } _ { t } } + \epsilon )$ , we expect that small values in the second state $\mathbf { r } _ { t }$ to produce large Adam updates. To get a better quantization error for small values we can switch the dynamic exponent and the base exponent. For regular dynamic quantization the base exponent is $1 0 ^ { 0 } \stackrel { } { = } 1$ and each zero bit decreases the exponent by a factor of 10 for a minimum value of $\mathrm { \dot { 1 } 0 ^ { - 7 } }$ . We invert this starting with base $1 0 ^ { - 7 }$ and each zero bit increases the exponent by 10 for a maximum value of 1. We denote this quantization as inverse dynamic quantization.
+
+# E.2 QUANTILE QUANTIZATION: A LOSSY MINIMUM ENTROPY ENCODING
+
+A lossy minimum entropy encoding with $k$ bits has the property that for any input data, the quantized outputs take the value of each of the $2 ^ { k }$ different bit representations equally often.
+
+More formally, a lossy minimum entropy encoding can be described in the following way. Given an infinite stream of sampled real numbers $x _ { i }$ where $x _ { i }$ is distributed as $X$ , an arbitrary probability distribution, a lossy minimum entropy encoding is given by the $k$ -bit quantization map $\mathbf { Q } ^ { \mathrm { m a p } } \in \mathbb { R } ^ { 2 ^ { k } }$ which maps values $q \in \mathbb { R } ^ { 2 ^ { k } }$ to indices $0 , 1 , \ldots 2 ^ { k }$ which has the property that if any number of elements $x _ { i }$ from the stream are quantized to $x _ { i } ^ { q }$ we do not gain any information which is predictive of future x j>i.
+
+One way to fulfill this property for arbitrary probability distributions $X$ , is to divide the probability distribution function $f _ { X }$ into $\bar { 2 } ^ { k }$ bins where each bin has equal area and the mid-points of these bins are values $q$ of the quantization map $\mathbf { Q } ^ { \mathrm { m a p } }$ . Empirically, this is equivalent to a histogram with $2 ^ { k }$ bins where each bin contains equal number of values.
+
+How do we find the mid-points for each histogram bin? This is equivalent to finding the $2 ^ { k }$ nonoverlapping values $x$ for the cumulative distribution function $F _ { X }$ with equal probability mass. These values can most easily be found by using its inverse function, the quantile function $\dot { Q _ { X } } = F _ { X } ^ { - 1 }$ . We can find the mid-points of each of the histogram bins by using the mid-points between $2 ^ { k } + 1$ equally spaced quantiles over the range of probabilities $[ 0 , 1 ]$ :
+
+$$
+q _ { i } = \frac { Q _ { X } \left( \frac { i } { 2 ^ { k } + 1 } \right) + Q _ { X } \left( \frac { i + 1 } { 2 ^ { k } + 1 } \right) } { 2 } ,
+$$
+
+To find $q$ empirically, we can estimate sample quantiles for a tensor $\mathbf { T }$ with unknown distribution $X$ by finding the $2 ^ { k }$ equally spaced sample quantiles via $\mathbf { T }$ ’s empirical cumulative distribution function. We refer to this quantization as quantile quantization.
+
+To estimate sample quantiles efficiently, we devise a specialized approximate quantile estimation algorithm, SRAM-Quantiles, which is more than $7 5 \mathrm { x }$ faster than other approximate quantile estimation approaches (Govindaraju et al., 2005; Dunning and Ertl, 2019). SRAM-Quantiles uses a divide-and-conquer strategy to perform sorting solely in fast SRAM. More details on this algorithm can be found in the Appendix Section F.
+
+# E.3 VISUALIZATION: DYNAMIC VS LINEAR QUANTIZATION VS QUANTILE QUANTIZATION
+
+Figure 6 shows the mapping from each to the 255 values of the 8-bit data types to their value normalized in the range [-1, 1]. We can see that most bits in dynamic quantization are allocated for large and small values. Quantile quantization is introduced in Appendix E.2.
+
+![](images/f0d98d3f97aad7ab5339ec027d769877c41d7f7028de27f5029c043996811292.jpg)  
+Figure 6: Visualization of the quantization maps for the linear, dynamic and quantile quantization. For quantile quantization we use values from the standard normal distribution and normalize them into the range [-1, 1].
+
+# F SRAM-QUANTILES: A FAST QUANTILE ESTIMATION ALGORITHM
+
+To estimate sample quantiles of a tensor one needs to determine the empirical cumulative distribution function (eCDF) of that tensor. The easiest way to find the eCDF is to sort a given tensor. Once sorted, the quantiles can be found by using the value at index $i = q \times n$ where $i$ is the index into the sorted array, $q$ is the desired quantile and $n$ is the total elements in the tensor. While simple, this process of estimating quantiles is computationally expensive and would render training with quantile quantization too slow to be useful.
+
+Similar to other quantile estimation approaches, our GPU algorithm, SRAM-Quantiles, uses a sliding windows over the data for fast, approximate quantile estimation with minimal resources. Greenwald and Khanna (2001)’s quantile estimation algorithm uses dynamic bin histograms over sliding windows to estimate quantiles. Extensions of this algorithm accelerate estimation by using more efficient data structures and estimation algorithms (Dunning and Ertl, 2019) or by using GPUs (Govindaraju et al., 2005). The main difference between this work an ours is that we only compute a limit set of quantiles that are known a priori – 256, to be exact – while previous work focuses on general statistics which help to produce any quantile a posteriori. Thus we can devise a highly specialized algorithm which offers faster estimation.
+
+The idea behind our algorithm comes from the fact that sorting is slow because it involves repeated loads and stores from main memory (DRAM) when executing divide-and-conquer sorting algorithms. We can significantly improve performance of quantile estimation if we restructure quantile estimation to respect memory hierarchies of the device on which the algorithm is executed.
+
+On a GPU, programmable SRAM – known as shared memory – is $1 5 \mathrm { x }$ faster than DRAM but has a limit size of around $6 4 \mathrm { k b }$ per core. The SRAM-Quantiles algorithm is simple. Instead of finding the full eCDF we find the eCDF for a subset of values of the tensor that fits into SRAM (about 4096 32-bit values). Once we found the quantiles for each subset, we average the quantiles atomically in DRAM.
+
+This algorithm works, because the arithmetic mean is an unbiased estimator for the population mean and samples quantiles estimated via eCDFs are asymptotically unbiased estimators of the population quantile (Chen and Kelton, 2001). Thus the more subset quantiles we average, the better the estimate of the tensor-wide quantiles.
+
+For estimating 256 quantiles on a large stream of numbers, our algorithm takes on average 0.064 ns to process one element in the stream, whereas the fastest general algorithms take 300 ns (Govindaraju et al., 2005) and 5 ns (Dunning and Ertl, 2019).
+
+# G ADAGRAD COMPARISONS
+
+While the main aim in this work is to investigate how the most commonly used optimizers, such as Adam (Kingma and Ba, 2014) and Momentum (Qian, 1999), can be used as 8-bit variants without any further hyperparameter tuning, it can be of interest to consider the behavior of our 8-bit methods under different scenarios. For example, one difference between Adam/Momentum and AdaGrad (Duchi et al., 2011) is that AdaGrad accumulates gradients statistics over the entire course of training while Adam/Momentum use a smoothed exponential decay over time. As such, this could lead to very different 8-bit quantization behavior where there are large difference between the magnitude of different optimizer states. Such large differences could induce a large quantization error and degrade performance of 8-bit optimizers.
+
+To investigate this, we train small 209M parameter language models on the RoBERTa corpus (Liu et al., 2019). We use the AdaGrad hyperparameters introduced by Keskar et al. (2019). Results are shown in Table 7. From the results we can see that our 8-bit methods do not work as well for AdaGrad. One hypothesis is that this is due to the the wide range of gradient statistics of AdaGrad which comes from averaging the gradient over the entire course of training. To prevent poor quantization in such scenarios, stochastic rounding proved to be very effective from our initial experiments with other 8-bit optimizer. While we abandoned stochastic rounding because we did not see any benefits for Adam and Momentum, it could be an effective solution for AdaGrad. We leave such improved 8-bit quantization methods for AdaGrad to future work.
+
+Table 7: AdaGrad compared to Adam performance for a 209M parameter language model on the RoBERTa corpus. The 8-bit methods use stable embedding layer. AdaGrad hyperparamters are taken from (Keskar et al., 2019).
+
+<table><tr><td>Optimizer</td><td>Valid Perplexity</td></tr><tr><td>32-bit Adam 8-bit Adam</td><td>16.7</td></tr><tr><td>32-bit AdaGrad</td><td>16.4 19.4</td></tr><tr><td>8-bit AdaGrad</td><td>19.7</td></tr><tr><td></td><td></td></tr></table>

parse/train/BJE-4xW0W/BJE-4xW0W_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0e12db665bb657323220419d9baab09f42343583b38dcdf787657dc5f6859ae
+size 6320452

parse/train/BJE-4xW0W/BJE-4xW0W_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92b7c289f2734fa6ea96c03f7e59a974b3bcbcc286c1571cff3a8e3822551fa4
+size 5939230

parse/train/BJE-4xW0W/BJE-4xW0W_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb11407801175f8e71d12dc0783dfd5e97347fe38c227b755960f96f444690aa
+size 6333080

parse/train/BJbD_Pqlg/BJbD_Pqlg_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d358dcb93c5ed70e69af0f0c355c1acf05c148b253523891edbfff70d6407a14
+size 2282625

parse/train/BJbD_Pqlg/BJbD_Pqlg_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2afa36a3fc8dec4827688b8e4c26c847e57e3e275797f5b88f36618b9dcbd9e
+size 2139820

parse/train/BJbD_Pqlg/BJbD_Pqlg_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f63b6cf33a007d9a968eb5d77fa6f03f71ca2e13664d9c6f75889a147a5577a
+size 2280767

parse/train/BJj6qGbRW/BJj6qGbRW_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88c677101b68559067b6551c8ed4b8afe70c6e6d36d8dde622161ba1383d3cdf
+size 1839614

parse/train/BJj6qGbRW/BJj6qGbRW_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb504b4f2aa078bd43a127b1538b2535046215b79c09b6f835595c2dc3863520
+size 1716829

parse/train/BJj6qGbRW/BJj6qGbRW_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c02b5c542a76d6b15f916af4afd08747fc8ccd5be6b58e8b12e256dc82141fc5
+size 1842495

parse/train/BJl_VnR9Km/BJl_VnR9Km_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3d775d8e0b5d7e99e4dbdc8c99e47ca8da1aa0dbe682b5785fb9f4e08e70ac6
+size 2215344

parse/train/BJl_VnR9Km/BJl_VnR9Km_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c16feb820d79c7d2d0b7ef1feeae1944b8f18378c831834e8e8e362200856bd2
+size 2067165

parse/train/BJl_VnR9Km/BJl_VnR9Km_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da98682b85440bc99d799f6d887309a9efcfd2ee4f76e00309971299313e0890
+size 2223049

parse/train/BkgnhTEtDS/BkgnhTEtDS_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3045596dbec7031652e065ec428608686fe3adee102de3aeae75ac3a73d02209
+size 2949229

parse/train/BkgnhTEtDS/BkgnhTEtDS_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67b17aa573f6f86a5074404e1c4a28e135de0a75399cee8caa27de091ed08dc9
+size 2737872

parse/train/BkgnhTEtDS/BkgnhTEtDS_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d23da1ee641ce7eb6b01405574e10bf1f08d400a002317b0a7c5a4925e91caf
+size 2955067

parse/train/ByG8A7cee/ByG8A7cee_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63bab47bf6e1f620050e5bbd23f889f9d467ca3208f001a6d719bff33eaf1200
+size 760905

parse/train/ByG8A7cee/ByG8A7cee_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc580530cd62a327631de3db2a28ab8b33922724cf42d4161602b23d23212396
+size 590065

parse/train/ByG8A7cee/ByG8A7cee_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0477969931c54d2d0e25b3e71c3550aa72ee9daea8e7930750208a230ce49e5
+size 764342

parse/train/CoJibBRjPXQ/CoJibBRjPXQ_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b34f2cf809cfe2803d1ad3cf1385b3866a4de9391a81ca8135f0b7e2d6745c4
+size 2325395

parse/train/CoJibBRjPXQ/CoJibBRjPXQ_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:594396a834da93e80cdd70aa3f7f0bc4f765b4c779d6e1ee918dfcb1e93d13e9
+size 2039507

parse/train/CoJibBRjPXQ/CoJibBRjPXQ_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c25eef1a52164d50e105e5fa4db64709265aa3c635dc06f51b8c33c48c03e227
+size 2362447

parse/train/H1e_cC4twS/H1e_cC4twS_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8dbeb19429fc9a8edc0d85328c4daa8078083284bfa35319ed6afd262668e862
+size 1565252

parse/train/H1e_cC4twS/H1e_cC4twS_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ce1dd5bfa47dc2123bbbb1fb9214a891438eb19ac828dd852912ddb7c44aad1
+size 1367762

parse/train/H1e_cC4twS/H1e_cC4twS_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e328628a80df6c7ba19abf2a66d89b3c1d2ccfd9000dfe47ca7a0a2a49effecd
+size 1569830

parse/train/H1eqOnNYDH/H1eqOnNYDH_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d26f57ab22b285fd683bbd1d6c47ebba9f7f8168858407071148aa77bdf9e87
+size 543200

parse/train/H1eqOnNYDH/H1eqOnNYDH_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16a4c7e2d5258340b5ab1f025b7d5bd6bde5c86508f373de37c7534dee5ecc47
+size 374324

parse/train/H1eqOnNYDH/H1eqOnNYDH_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:228c34e764bcda1318e783a1f1ecd44590d97fadf0af79e72576adc8a22ea1ff
+size 544877

parse/train/H1xk8jAqKQ/H1xk8jAqKQ_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a85ee0bb5830ca90de58a1676a36ffe12f03f40ec05170b6158faa65ea85f3fc
+size 1999407

parse/train/H1xk8jAqKQ/H1xk8jAqKQ_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46f07ad7558eaa841c4a8db7e300d776c4e9fc673d8eaf876e3dbc3a341bb1de
+size 1824253

parse/train/H1xk8jAqKQ/H1xk8jAqKQ_span.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a459cb102555cf2ebccd4afba32cdbcc39690f7121d50c4d74e6429cffd57231
+size 2001128

parse/train/HJBhEMbRb/HJBhEMbRb_layout.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd7ae0e615fce55bd7ba58c939b8a3f8991998a8b7501995679d6f1655e8c004
+size 1396179