We introduce Perception Encoder (PE), a family of state-of-the-art vision encoders for image and video understanding. Traditionally, vision encoders have relied on a variety of pretraining objectives, each excelling at different downstream tasks. Surprisingly, after scaling a carefully tuned image pretraining recipe and refining with a robust video data engine, we find that contrastive vision-language training alone can produce strong, general embeddings for all of these downstream tasks. There is only one caveat: these embeddings are hidden within the intermediate layers of the network. To draw them out, we introduce two alignment methods: language alignment for multimodal language modeling, and spatial alignment for dense prediction. Together, our PE family of models achieves state-of-the-art results on a wide variety of tasks, including zero-shot image and video classification and retrieval; document, image, and video Q&A; and spatial tasks such as detection, tracking, and depth estimation. We release our models, code, and novel dataset of synthetically and human-annotated videos: https://github.com/facebookresearch/perception_models

Vision-language models are integral to computer vision research, yet many high-performing models remain closed-source, obscuring their data, design and training recipe. The research community has responded by using distillation from black-box models to label training data, achieving strong benchmark results, at the cost of measurable scientific progress. However, without knowing the details of the teacher model and its data sources, scientific progress remains difficult to measure. In this paper, we study building a Perception Language Model (PLM) in a fully open and reproducible framework for transparent research in image and video understanding. We analyze standard training pipelines without distillation from proprietary models and explore large-scale synthetic data to identify critical data gaps, particularly in detailed video understanding. To bridge these gaps, we release 2.8M human-labeled instances of fine-grained video question-answer pairs and spatio-temporally grounded video captions. Additionally, we introduce PLM–VideoBench, a suite for evaluating challenging video understanding tasks focusing on the ability to reason about ''what'', ''where'', ''when'', and ''how'' of a video. We make our work fully reproducible by providing data, training recipes, code & models.

We introduce VOICECRAFT, a token infilling neural codec language model, that achieves state-of-the-art performance on both speech editing and zero-shot text-to-speech (TTS) on audiobooks, internet videos, and podcasts 1 .VOICECRAFT employs a Transformer decoder architecture and introduces a token rearrangement procedure that combines causal masking and delayed stacking to enable generation within an existing sequence.On speech editing tasks, VOICECRAFT produces edited speech that is nearly indistinguishable from unedited recordings in terms of naturalness, as evaluated by humans; for zero-shot TTS, our model outperforms prior SotA models including VALL-E and the popular commercial model XTTS v2.Crucially, the models are evaluated on challenging and realistic datasets, that consist of diverse accents, speaking styles, recording conditions, and background noise and music, and our model performs consistently well compared to other models and real recordings.In particular, for speech editing evaluation, we introduce a high quality, challenging, and realistic dataset named REALEDIT.We encourage readers to listen to the demos at https: //jasonppy.github.io/VoiceCraft_web.

The success of contrastive language-image pretraining (CLIP) relies on the supervision from the pairing between images and captions, which tends to be noisy in web- crawled data. We present Mixture of Data Experts (MoDE) and learn a system of CLIP data experts via clustering. Each data expert is trained on one data cluster, being less sensitive to false negative noises in other clusters. At inference time, we ensemble their outputs by applying weights determined through the correlation between task metadata and cluster conditions. To estimate the correlation pre-cisely, the samples in one cluster should be semantically similar, but the number of data experts should still be rea-sonable for training and inference. As such, we consider the ontology in human language and propose to use fine- grained cluster centers to represent each data expert at a coarse-grained level. Experimental studies show that four CLIP data experts on ViT-B/16 outperform the ViT-L/14 by OpenAI CLIP and OpenCLIP on zero-shot image classification but with less (<35%) training cost. Meanwhile, MoDE can train all data expert asynchronously and can flexibly include new data experts. The code is available here.

Hu Xu, Po-Yao Huang, Xiaoqing Tan, Ching-Feng Yeh, Jacob Kahn, Christine Jou, Gargi Ghosh, Omer Levy, Luke Zettlemoyer, Wen-tau Yih, Shang-Wen Li, Saining Xie, Christoph Feichtenhofer. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Audio-visual representation learning aims to develop systems with human-like perception by utilizing correlation between auditory and visual information. However, current models often focus on a limited set of tasks, and generalization abilities of learned representations are unclear. To this end, we propose the AV-SUPERB benchmark that enables general-purpose evaluation of unimodal audio/visual and bimodal fusion representations on 7 datasets covering 5 audio-visual tasks in speech and audio processing. We evaluate 5 recent self-supervised models and show that none of these models generalize to all tasks, emphasizing the need for future study on improving universal model performance. In addition, we show that representations may be improved with intermediate-task fine-tuning and audio event classification with AudioSet serves as a strong intermediate task. We release our benchmark with evaluation code1and a model submission platform2to encourage further research in audio-visual learning.

Contrastive Language-Image Pre-training (CLIP) is an approach that has advanced research and applications in computer vision, fueling modern recognition systems and generative models. We believe that the main ingredient to the success of CLIP is its \textit{data} and \textit{not} the \textit{model} architecture or pre-training {objective}. However, CLIP only provides very limited information about its data and how it has been collected, leading to works that aim to reproduce CLIP's data by filtering with its model parameters. In this work, we intend to reveal CLIP's data curation approach and in our pursuit of making it open to the community introduce Metadata-Curated Language-Image Pre-training (MetaCLIP). MetaCLIP takes a raw data pool and metadata (derived from CLIP's concepts) and yields a balanced subset over the metadata distribution. Our experimental study rigorously isolates the model and training settings, concentrating solely on data. MetaCLIP applied to CommonCrawl with 400M image-text data pairs outperforms CLIP's data on multiple standard benchmarks. In zero-shot ImageNet classification, MetaCLIP achieves 70.8\% accuracy, surpassing CLIP's 68.3\% on \mbox{ViT-B} models. Scaling to 1B data, while maintaining the same training budget, attains \textbf{72.4\%}. Our observations hold across various model sizes, exemplified by ViT-H achieving \textbf{80.5\%}, without any bells-and-whistles. Curation code and training data distribution over metadata will be made available.

Tiezheng Yu, Hanchao Yu, Davis Liang, Yuning Mao, Shaoliang Nie, Po-Yao Huang, Madian Khabsa, Pascale Fung, Yi-Chia Wang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

We propose Fast Language-Audio Pre-training (FLAP), a self-supervised approach that efficiently and effectively learns aligned audio and language representations through masking, contrastive learning and reconstruction. For efficiency, FLAP randomly drops audio spectrogram tokens, focusing solely on the remaining ones for self-supervision. Through inter-modal contrastive learning, FLAP learns to align paired audio and text representations in a shared latent space. Notably, FLAP leverages multiple augmented views via masking for intermodal contrast and learns to reconstruct the masked portion of audio tokens. Moreover, FLAP leverages large language models (LLMs) to augment the text inputs, contributing to improved performance. These approaches lead to more robust and informative audio-text representations, enabling FLAP to achieve state-of-the-art (SoTA) performance on audio-text retrieval tasks on AudioCaps (achieving 53.0% R@1) and Clotho (achieving 25.5% R@1).

We study the problem of human action recognition using motion capture (MoCap) sequences. Unlike existing techniques that take multiple manual steps to derive standard-ized skeleton representations as model input, we propose a novel Spatial-Temporal Mesh Transformer (STMT) to directly model the mesh sequences. The model uses a hierarchical transformer with intra-frame off-set attention and inter-frame self-attention. The attention mechanism allows the model to freely attend between any two vertex patches to learn nonlocal relationships in the spatial-temporal domain. Masked vertex modeling and future frame prediction are used as two self-supervised tasks to fully activate the bi-directional and auto-regressive attention in our hierarchical transformer. The proposed method achieves state-of-the-art performance compared to skeleton-based and point-cloud-based models on common MoCap benchmarks. Code is available at https://github.com/zgzxy001/STMT.

Large vision-language models are generally applicable to many downstream tasks, but come at an exorbitant training cost that only large institutions can afford. This paper trades generality for efficiency and presents Curation in Training (CiT), a simple and efficient vision-text learning algorithm that couples a data objective into training. CiT automatically yields quality data to speed-up contrastive image-text training and alleviates the need for an offline data filtering pipeline, allowing broad data sources (including raw image-text pairs from the web). CiT contains two loops: an outer loop curating the training data and an inner loop consuming the curated training data. The text encoder connects the two loops. Given metadata for tasks of interest, e.g., class names, and a large pool of image-text pairs, CiT alternatively selects relevant training data from the pool by measuring the similarity of their text embeddings and embeddings of the metadata. In our experiments, we observe that CiT can speed up training by over an order of magnitude, especially if the raw data size is large.

There has been a longstanding belief that generation can facilitate a true understanding of visual data. In line with this, we revisit generatively pre-training visual representations in light of recent interest in denoising diffusion models. While directly pre-training with diffusion models does not produce strong representations, we condition diffusion models on masked input and formulate diffusion models as masked autoencoders (DiffMAE). Our approach is capable of (i) serving as a strong initialization for downstream recognition tasks, (ii) conducting high-quality image inpainting, and (iii) being effortlessly extended to video where it produces state-of-the-art classification accuracy. We further perform a comprehensive study on the pros and cons of design choices and build connections between diffusion models and masked autoencoders. Project page.

Modern hierarchical vision transformers have added several vision-specific components in the pursuit of supervised classification performance. While these components lead to effective accuracies and attractive FLOP counts, the added complexity actually makes these transformers slower than their vanilla ViT counterparts. In this paper, we argue that this additional bulk is unnecessary. By pretraining with a strong visual pretext task (MAE), we can strip out all the bells-and-whistles from a state-of-the-art multi-stage vision transformer without losing accuracy. In the process, we create Hiera, an extremely simple hierarchical vision transformer that is more accurate than previous models while being significantly faster both at inference and during training. We evaluate Hiera on a variety of tasks for image and video recognition. Our code and models are available at https://github.com/facebookresearch/hiera.

We present Masked Audio-Video Learners (MAViL) to learn audio-visual representations with three complementary forms of self-supervision: (1) reconstructing masked raw audio and video inputs, (2) intra-modal and inter-modal contrastive learning with masking, and (3) self-training to predict aligned and contextualized audio-video representations learned from the first two objectives. Empirically, MAViL achieves state-of-the-art audio-video classification performance on AudioSet (53.3 mAP) and VGGSound (67.1\% accuracy), surpassing recent self-supervised models and supervised models that utilize external labeled data. Notably, pre-training with MAViL not only enhances performance in multimodal classification and retrieval tasks, but it also improves the representations of each modality in isolation, without relying on information from the other modality during uni-modal fine-tuning or inference. The code and models are available at https://github.com/facebookresearch/MAViL.

The main challenge in the field of unsupervised machine translation (UMT) is to associate source-target sentences in the latent space. As people who speak different languages share biologically similar visual systems, various unsupervised multi-modal machine translation (UMMT) models have been proposed to improve the performances of UMT by employing visual contents in natural images to facilitate alignment. Commonly, relation information is the important semantic in a sentence. Compared with images, videos can better present the interactions between objects and the ways in which an object transforms over time. However, current state-of-the-art methods only explore scene-level or object-level information from images without explicitly modeling objects relation; thus, they are sensitive to spurious correlations, which poses a new challenge for UMMT models. In this paper, we employ a spatial-temporal graph obtained from videos to exploit object interactions in space and time for disambiguation purposes and to promote latent space alignment in UMMT. Our model employs multi-modal back-translation and features pseudo-visual pivoting, in which we learn a shared multilingual visual-semantic embedding space and incorporate visually pivoted captioning as additional weak supervision. Experimental results on the VATEX Translation 2020 and HowToWorld datasets validate the translation capabilities of our model on both sentence-level and word-level and generalizes well when videos are not available during the testing phase.

As audio-visual systems are being deployed for safety-critical tasks such as surveillance and malicious content filtering, their robustness remains an under-studied area. Existing published work on robustness either does not scale to large-scale dataset, or does not deal with multiple modalities. This work aims to study several key questions related to multi-modal learning through the lens of robustness: 1) Are multi-modal models necessarily more robust than uni-modal models? 2) How to efficiently measure the robustness of multi-modal learning? 3) How to fuse different modalities to achieve a more robust multi-modal model? To understand the robustness of the multi-modal model in a large-scale setting, we propose a density-based metric, and a convexity metric to efficiently measure the distribution of each modality in high-dimensional latent space. Our work provides a theoretical intuition together with empirical evidence showing how multi-modal fusion affects adversarial robustness through these metrics. We further devise a mix-up strategy based on our metrics to improve the robustness of the trained model. Our experiments on AudioSet [1] and Kinetics-Sounds [2] verify our hypothesis that multi-modal models are not necessarily more robust than their uni-modal counterparts in the face of adversarial examples. We also observe our mix-up trained method could achieve as much protection as traditional adversarial training, offering a computationally cheap alternative.

After its sweeping success in vision and language tasks, pure attention-based neural architectures (e.g.DeiT) [1] are emerging to the top of audio tagging (AT) leaderboards [2], which seemingly obsoletes traditional convolutional neural networks (CNNs), feed-forward networks or recurrent networks.However, taking a closer look, there is great variability in published research, for instance, performances of models initialized with pretrained weights differ drastically from without pretraining [2], training time for a model varies from hours to weeks, and often, essences are hidden in seemingly trivial details.This urgently calls for a comprehensive study since our 1st comparison [3] is half-decade old.In this work, we perform extensive experiments on AudioSet [4] which is the largest weakly-labeled sound event dataset available, we also did analysis based on the data quality and efficiency.We compare a few state-of-the-art baselines on the AT task, and study the performance and efficiency of 2 major categories of neural architectures: CNN variants and attention-based variants.We also closely examine their optimization procedures.Our opensourced experimental results 1 provide insights to trade off between performance, efficiency, optimization process, for both practitioners and researchers. 2

This paper studies a simple extension of image-based Masked Autoencoders (MAE) to self-supervised representation learning from audio spectrograms. Following the Transformer encoder-decoder design in MAE, our Audio-MAE first encodes audio spectrogram patches with a high masking ratio, feeding only the non-masked tokens through encoder layers. The decoder then re-orders and decodes the encoded context padded with mask tokens, in order to reconstruct the input spectrogram. We find it beneficial to incorporate local window attention in the decoder, as audio spectrograms are highly correlated in local time and frequency bands. We then fine-tune the encoder with a lower masking ratio on target datasets. Empirically, Audio-MAE sets new state-of-the-art performance on six audio and speech classification tasks, outperforming other recent models that use external supervised pre-training. Our code and models is available at https://github.com/facebookresearch/AudioMAE.

Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, Christoph Feichtenhofer. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

As audio/visual classification models are widely deployed for sensitive tasks like content filtering at scale, it is critical to understand their robustness along with improving the accuracy. This work aims to study several key questions related to multimodal learning through the lens of adversarial noises: 1) The trade-off between early/middle/late fusion affecting its robustness and accuracy 2) How does different frequency/time domain features contribute to the robustness? 3) How does different neural modules contribute against the adversarial noise? In our experiment, we construct adversarial examples to attack state-of-the-art neural models trained on Google AudioSet.[1]1We compare how much attack potency in terms of adversarial perturbation of size using different Lpnorms we would need to "deactivate" the victim model. Using adversarial noise to ablate multimodal models, we are able to provide insights into what is the best potential fusion strategy to balance the model parameters/accuracy and robustness trade-off, and distinguish the robust features versus the non-robust features that various neural networks model tend to learn.

The quality of the image representations obtained from self-supervised learning depends strongly on the type of data augmentations used in the learning formulation. Recent papers have ported these methods from still images to videos and found that leveraging both audio and video signals yields strong gains; however, they did not find that spatial augmentations such as cropping, which are very important for still images, work as well for videos. In this paper, we improve these formulations in two ways unique to the spatio-temporal aspect of videos. First, for space, we show that spatial augmentations such as cropping do work well for videos too, but that previous implementations, due to the high processing and memory cost, could not do this at a scale sufficient for it to work well. To address this issue, we first introduce Feature Crop, a method to simulate such augmentations much more efficiently directly in feature space. Second, we show that as opposed to naïve average pooling, the use of transformer-based attention improves performance significantly, and is well suited for processing feature crops. Combining both of our discoveries into a new method, Space-Time Crop & Attend (STiCA) we achieve state-of-the-art performance across multiple video-representation learning benchmarks. In particular, we achieve new state-of-the-art accuracies of 67.0% on HMDB-51 and 93.1% on UCF-101 when pre-training on Kinetics-400. Code and pretrained models are available1.

Po-Yao Huang, Mandela Patrick, Junjie Hu, Graham Neubig, Florian Metze, Alexander Hauptmann. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

The training of a feature extraction network typically requires abundant manually annotated training samples, making this a time-consuming and costly process. Accordingly, we propose an effective self-supervised learning-based tracker in a deep correlation framework (named: self-SDCT). Motivated by the forward-backward tracking consistency of a robust tracker, we propose a multi-cycle consistency loss as self-supervised information for learning feature extraction network from adjacent video frames. At the training stage, we generate pseudo-labels of consecutive video frames by forward-backward prediction under a Siamese correlation tracking framework and utilize the proposed multi-cycle consistency loss to learn a feature extraction network. Furthermore, we propose a similarity dropout strategy to enable some low-quality training sample pairs to be dropped and also adopt a cycle trajectory consistency loss in each sample pair to improve the training loss function. At the tracking stage, we employ the pre-trained feature extraction network to extract features and utilize a Siamese correlation tracking framework to locate the target using forward tracking alone. Extensive experimental results indicate that the proposed self-supervised deep correlation tracker (self-SDCT) achieves competitive tracking performance contrasted to state-of-the-art supervised and unsupervised tracking methods on standard evaluation benchmarks.

Unsupervised machine translation (MT) has recently achieved impressive results with monolingual corpora only.However, it is still challenging to associate source-target sentences in the latent space.As people speak different languages biologically share similar visual systems, the potential of achieving better alignment through visual content is promising yet under-explored in unsupervised multimodal MT (MMT).In this paper, we investigate how to utilize visual content for disambiguation and promoting latent space alignment in unsupervised MMT.Our model employs multimodal back-translation and features pseudo visual pivoting in which we learn a shared multilingual visual-semantic embedding space and incorporate visuallypivoted captioning as additional weak supervision.The experimental results on the widely used Multi30K dataset show that the proposed model significantly improves over the state-ofthe-art methods and generalizes well when images are not available at the testing time.

We explore methods to enrich the diversity of captions associated with pictures for learning improved visual-semantic embeddings (VSE) in cross-modal retrieval. In the spirit of "A picture is worth a thousand words", it would take dozens of sentences to parallel each picture's content adequately. But in fact, real-world multimodal datasets tend to provide only a few (typically, five) descriptions per image. For cross-modal retrieval, the resulting lack of diversity and coverage prevents systems from capturing the fine-grained inter-modal dependencies and intra-modal diversities in the shared VSE space. Using the fact that the encoder-decoder architectures in neural machine translation (NMT) have the capacity to enrich both monolingual and multilingual textual diversity, we propose a novel framework leveraging multimodal neural machine translation (MMT) to perform forward and backward translations based on salient visual objects to generate additional text-image pairs which enables training improved monolingual cross-modal retrieval (English-Image) and multilingual cross-modal retrieval (English-Image and German-Image) models. Experimental results show that the proposed framework can substantially and consistently improve the performance of state-of-the-art models on multiple datasets. The results also suggest that the models with multilingual VSE outperform the models with monolingual VSE.

Po-Yao Huang, Xiaojun Chang, Alexander Hauptmann. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Although significant progress has been made for cross-modal retrieval models in recent years, few have explored what those models truly learn and what makes one model superior to another. Start by training two state-of-the-art text-to-image retrieval models with adversarial text inputs, we investigate and quantify the importance of syntactic structure and lexical information in learning the joint visual-semantic embedding space for cross-modal retrieval. The results show that the retrieval power mainly comes from localizing and connecting the visual objects and their cross-modal counter-parts, the textual phrases. Inspired by this observation, we propose a novel model which employs object-oriented encoders along with inter- and intra-modal attention networks to improve inter-modal dependencies for cross-modal retrieval. In addition, we develop a new multimodal structure-preserving objective which additionally emphasizes intra-modal hard negative examples to promote intra-modal discrepancies. Extensive experiments show that the proposed approach outperforms the existing best method by a large margin (16.4% and 6.7% relatively with [email protected] in the text-to-image retrieval task on the Flickr30K dataset and the MS-COCO dataset respectively).

Visual-semantic embeddings are central to many multimedia applications such as cross-modal retrieval between visual data and natural language descriptions. Conventionally, learning a joint embedding space relies on large parallel multimodal corpora. Since massive human annotation is expensive to obtain, there is a strong motivation in developing versatile algorithms to learn from large corpora with fewer annotations. In this paper, we propose a novel framework to leverage automatically extracted regional semantics from un-annotated images as additional weak supervision to learn visual-semantic embeddings. The proposed model employs adversarial attentive alignments to close the inherent heterogeneous gaps between annotated and un-annotated portions of visual and textual domains. To demonstrate its superiority, we conduct extensive experiments on sparsely annotated multimodal corpora. The experimental results show that the proposed model outperforms state-of-the-art visual-semantic embedding models by a significant margin for cross-modal retrieval tasks on the sparse Flickr30k and MS-COCO datasets. It is also worth noting that, despite using only 20% of the annotations, the proposed model can achieve competitive performance (Recall at 10 > 80.0% for 1K and > 70.0% for 5K text-to-image retrieval) compared to the benchmarks trained with the complete annotations.

Developing an efficient and effective social media monitoring system has become one of the important steps towards improved public safety. With the explosive availability of user-generated content documenting most conflicts and human rights abuses around the world, analysts and first-responders increasingly find themselves overwhelmed with massive amounts of noisy data from social media. In this paper, we construct a large-scale public safety event dataset with retrospective automatic labeling for 4.2 million multimodal tweets from 7 public safety events occurred in 2013~2017. We propose a new multimodal social media filtering system composed of encoding, classification, and correlation networks to jointly learn shared and complementary visual and textual information to filter out the most relevant and useful items among the noisy social media influx. The proposed model is verified and achieves significant improvement over competitive baselines under the retrospective and real-time experimental protocols.

What happened during the Boston Marathon in 2013? Nowadays, at any major event, lots of people take videos and share them on social media. To fully understand exactly what happened in these major events, researchers and analysts often have to examine thousands of these videos manually. To reduce this manual effort, we present an investigative system that automatically synchronizes these videos to a global timeline and localizes them on a map. In addition to alignment in time and space, our system combines various functions for analysis, including gunshot detection, crowd size estimation, 3D reconstruction and person tracking. To our best knowledge, this is the first time a unified framework has been built for comprehensive event reconstruction for social media videos.

In the era of social media, a large number of user-generated videos are uploaded to the Internet every day, capturing events all over the world. Reconstructing the event truth based on information mined from these videos has been an emerging challenging task. Temporal alignment of videos “in the wild” which capture different moments at different positions with different perspectives is the critical step. In this paper, we propose a hierarchical approach to synchronize videos. Our system utilizes clustered audio-signatures to align video pairs. Global alignment for all videos is then achieved via forming alignable video groups with self-paced learning. Experiments on the Boston Marathon dataset show that the proposed method achieves excellent precision and robustness.

Zhiting Hu, Poyao Huang, Yuntian Deng, Yingkai Gao, Eric Xing. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

The cognitive capability of secondary users in multichannel wireless networks enables the functionalities of system parameter estimation and learning, so that a more intelligent channel access without interfering the primary users is possible. This paper proposes a novel cognitive channel access algorithm with threshold policy on the basis of a continuous-time Markov chain built by the estimated parameters. The secondary users could access the channel in a more intelligent fashion and thus the better quality of service can be achieved. The numerical results show that cognitive channel access can significantly increase the total utility of system while keeping blocking probability of primary users' requests under a predefined constraint.