The rapid advancements in large foundation models have significantly transformed the field of visual content generation, impacting domains such as image synthesis, video generation, and 3D modeling. This tutorial will provide an in-depth exploration of the state-of-the-art techniques and methodologies used in visual content generation, emphasizing the role of large-scale generative models. The tutorial will cover fundamental principles, model architectures, recent breakthroughs, and practical applications. We will discuss various generative paradigms, including diffusion models, autoregressive models, and large multimodal models, highlighting their strengths and limitations. Additionally, we will delve into the challenges of controllability, personalization, and realism in generated content, along with open research problems and future directions. By the end of the tutorial, attendees will gain a comprehensive understanding of contemporary visual content generation techniques and their applications, equipping them with the knowledge to leverage these models in their research and projects.

While text-to-video generation shows state-of-the-art results, fine-grained output control remains challenging for users relying solely on natural language prompts. In this work, we present FACTOR for fine-grained controllable video generation. FACTOR provides an intuitive interface where users can manipulate the trajectory and appearance of individual objects in conjunction with a text prompt. We propose a unified framework to integrate these control signals into an existing text-to-video model. Our approach involves a multimodal condition module with a joint encoder, control-attention layers, and an appearance augmentation mechanism. This design enables FACTOR to generate videos that closely align with detailed user specifications. Extensive experiments on standard benchmarks and user-provided inputs demonstrate a notable improvement in controllability by FACTOR over competitive baselines.

Language has emerged as a natural interface for image editing. In this paper, we introduce a method for region- based image editing driven by textual prompts, without the need for user-provided masks or sketches. Specifically, our approach leverages an existing pre-trained text-to-image model and introduces a bounding box generator to iden-tify the editing regions that are aligned with the textual prompts. We show that this simple approach enables flex-ible editing that is compatible with current image generation models, and is able to handle complex prompts featuring multiple objects, complex sentences, or lengthy para- graphs. We conduct an extensive user study to compare our method against state-of-the-art methods. The experiments demonstrate the competitive performance of our method in manipulating images with high fidelity and realism that correspond to the provided language descriptions. Our project webpage can be found at: https://yuanzelin.me/LearnableRegions_page.

While Large Language Models (LLMs) are the dominant models for generative tasks in language, they do not perform as well as diffusion models on image and video generation. To effectively use LLMs for visual generation, one crucial component is the visual tokenizer that maps pixel-space inputs to discrete tokens appropriate for LLM learning. In this paper, we introduce \modelname{}, a video tokenizer designed to generate concise and expressive tokens for both videos and images using a common token vocabulary. Equipped with this new tokenizer, we show that LLMs outperform diffusion models on standard image and video generation benchmarks including ImageNet and Kinetics. In addition, we demonstrate that our tokenizer surpasses the previously top-performing video tokenizer on two more tasks: (1) video compression comparable to the next-generation video codec (VCC) according to human evaluations, and (2) learning effective representations for action recognition tasks.

We present VideoPoet, a language model capable of synthesizing high-quality video from a large variety of conditioning signals. VideoPoet employs a decoder-only transformer architecture that processes multimodal inputs – including images, videos, text, and audio. The training protocol follows that of Large Language Models (LLMs), consisting of two stages: pretraining and task-specific adaptation. During pretraining, VideoPoet incorporates a mixture of multimodal generative objectives within an autoregressive Transformer framework. The pretrained LLM serves as a foundation that can be adapted for a range of video generation tasks. We present empirical results demonstrating the model’s state-of-the-art capabilities in zero-shot video generation, specifically highlighting the ability to generate high-fidelity motions. Project page: http://sites.research.google/videopoet/

Training diffusion models for audiovisual sequences allows for a range of generation tasks by learning conditional distributions of various input-output combinations of the two modalities. Nevertheless, this strategy often requires training a separate model for each task which is expensive. Here, we propose a novel training approach to effectively learn arbitrary conditional distributions in the audiovisual space. Our key contribution lies in how we parameterize the diffusion timestep in the forward diffusion process. Instead of the standard fixed diffusion timestep, we propose applying variable diffusion timesteps across the temporal dimension and across modalities of the inputs. This formulation offers flexibility to introduce variable noise levels for various portions of the input, hence the term mixture of noise levels. We propose a transformer-based audiovisual latent diffusion model and show that it can be trained in a task-agnostic fashion using our approach to enable a variety of audiovisual generation tasks at inference time. Experiments demonstrate the versatility of our method in tackling cross-modal and multimodal interpolation tasks in the audiovisual space. Notably, our proposed approach surpasses baselines in generating temporally and perceptually consistent samples conditioned on the input. Project page: neurips13025.github.io

Learning generative image models from various domains efficiently needs transferring knowledge from an image synthesis model trained on a large dataset. We present a recipe for learning vision transformers by generative knowledge transfer. We base our framework on generative vision transformers representing an image as a sequence of visual tokens with the autoregressive or non-autoregressive transformers. To adapt to a new domain, we employ prompt tuning, which prepends learnable tokens called prompts to the image token sequence and introduces a new prompt design for our task. We study on a variety of visual domains with varying amounts of training images. We show the effectiveness of knowledge transfer and a significantly better image generation quality.11https://github.com/google-research/generative_transfer

We introduce the MAsked Generative VIdeo Transformer, MAGVIT, to tackle various video synthesis tasks with a single model. We introduce a 3D tokenizer to quantize a video into spatial-temporal visual tokens and propose an embedding method for masked video token modeling to facilitate multi-task learning. We conduct extensive experiments to demonstrate the quality, efficiency, and flexibility of MAGVIT. Our experiments show that (i) MAGVIT performs favorably against state-of-the-art approaches and establishes the best-published FVD on three video generation benchmarks, including the challenging Kinetics-600. (ii) MAGVIT outperforms existing methods in inference time by two orders of magnitude against diffusion models and by 60x against autoregressive models. (iii) A single MAGVIT model supports ten diverse generation tasks and generalizes across videos from different visual domains. The source code and trained models will be released to the public at https://magvit.cs.cmu.edu.

We present Muse, a text-to-image Transformermodel that achieves state-of-the-art image genera-tion performance while being significantly moreefficient than diffusion or autoregressive models.Muse is trained on a masked modeling task indiscrete token space: given the text embeddingextracted from a pre-trained large language model(LLM), Muse learns to predict randomly maskedimage tokens. Compared to pixel-space diffusionmodels, such as Imagen and DALL-E 2, Muse issignificantly more efficient due to the use of dis-crete tokens and requires fewer sampling itera-tions; compared to autoregressive models such asParti, Muse is more efficient due to the use of par-allel decoding. The use of a pre-trained LLM en-ables fine-grained language understanding, whichtranslates to high-fidelity image generation andthe understanding of visual concepts such as ob-jects, their spatial relationships, pose, cardinalityetc. Our 900M parameter model achieves a newSOTA on CC3M, with an FID score of 6.06. TheMuse 3B parameter model achieves an FID of7.88 on zero-shot COCO evaluation, along with aCLIP score of 0.32. Muse also directly enables anumber of image editing applications without theneed to fine-tune or invert the model: inpainting,outpainting, and mask-free editing. More resultsand videos demonstrating editing are available at https://muse-icml.github.io/

Pre-trained large text-to-image models synthesize impressive images with an appropriate use of text prompts. However, ambiguities inherent in natural language, and out-of-distribution effects make it hard to synthesize arbitrary image styles, leveraging a specific design pattern, texture or material. In this paper, we introduce *StyleDrop*, a method that enables the synthesis of images that faithfully follow a specific style using a text-to-image model. StyleDrop is extremely versatile and captures nuances and details of a user-provided style, such as color schemes, shading, design patterns, and local and global effects. StyleDrop works by efficiently learning a new style by fine-tuning very few trainable parameters (less than 1\% of total model parameters), and improving the quality via iterative training with either human or automated feedback. Better yet, StyleDrop is able to deliver impressive results even when the user supplies only a *single* image specifying the desired style. An extensive study shows that, for the task of style tuning text-to-image models, StyleDrop on Muse convincingly outperforms other methods, including DreamBooth and textual inversion on Imagen or Stable Diffusion. More results are available at our project website: [https://styledrop.github.io](https://styledrop.github.io).

In this work, we introduce Semantic Pyramid AutoEncoder (SPAE) for enabling frozen LLMs to perform both understanding and generation tasks involving non-linguistic modalities such as images or videos. SPAE converts between raw pixels and interpretable lexical tokens (or words) extracted from the LLM's vocabulary. The resulting tokens capture both the rich semantic meaning and the fine-grained details needed for visual reconstruction, effectively translating the visual content into a language comprehensible to the LLM, and empowering it to perform a wide array of multimodal tasks. Our approach is validated through in-context learning experiments with frozen PaLM 2 and GPT 3.5 on a diverse set of image understanding and generation tasks. Our method marks the first successful attempt to enable a frozen LLM to generate image content while surpassing state-of-the-art performance in image understanding tasks, under the same setting, by over 25%.

Image captioning aims at automatically describing images by sentences. It often requires lots of paired image-sentence data for training. However, trained captioning models can hardly be applied to new domains in which some novel words exist. In this paper, we introduce the zero-shot novel object captioning task, where the machine generates descriptions about novel objects without extra training sentences. To tackle the challenging task, we mimic the way that babies talk about something unknown, i.e., using the word of a similar known object. Following this motivation, we build a key-value object memory by detection models, containing visual information and corresponding words for objects in the image. For those novel objects, we use words of most similar seen objects as proxy visual words to solve the out-of-vocabulary issue. We then propose a Switchable LSTM that incorporates knowledge from the object memory into sentence generation. The model has two switchable working modes, i.e., 1) generating the sentences like standard LSTMs and 2) retrieving proper nouns from the key-value memory. Thus our model is learned to fully disentangle language generation from training objects, and requires zero training sentence in describing novel objects. Experiments on three large-scale datasets demonstrate the ability of our method to describe novel concepts.

Generative transformers have experienced rapid popularity growth in the computer vision community in synthesizing high-fidelity and high-resolution images. The best generative transformer models so far, however, still treat an image naively as a sequence of tokens, and decode an image sequentially following the raster scan ordering (i.e. line-by-line). We find this strategy neither optimal nor efficient. This paper proposes a novel image synthesis paradigm using a bidirectional transformer decoder, which we term MaskGIT. During training, MaskGIT learns to predict randomly masked tokens by attending to tokens in all directions. At inference time, the model begins with generating all tokens of an image simultaneously, and then refines the image iteratively conditioned on the previous generation. Our experiments demonstrate that MaskGIT significantly outperforms the state-of-the-art transformer model on the ImageNet dataset, and accelerates autoregressive decoding by up to 48x. Besides, we illustrate that MaskGIT can be easily extended to various image editing tasks, such as inpainting, extrapolation, and image manipulation. Project page: masked-generative-image-transformer.github.io.

Aggressive data augmentation is a key component of the strong generalization capabilities of Vision Transformer (ViT). One such data augmentation technique is adversarial training (AT); however, many prior works [28,45] have shown that this often results in poor clean accuracy. In this work, we present pyramid adversarial training (PyramidAT), a simple and effective technique to improve ViT's overall performance. We pair it with a “matched” Dropout and stochastic depth regularization, which adopts the same Dropout and stochastic depth configuration for the clean and adversarial samples. Similar to the improvements on CNNs by AdvProp [61] (not directly applicable to ViT), our pyramid adversarial training breaks the trade-off between in-distribution accuracy and out-of-distribution robustness for ViT and related architectures. It leads to 1.82% absolute improvement on ImageNet clean accuracy for the ViT-B model when trained only on ImageNet-1K data, while simultaneously boosting performance on 7 ImageNet ro-bustness metrics, by absolute numbers ranging from 1.76% to 15.68%. We set a new state-of-the-art for ImageNet-C (41.42 mCE), ImageNet-R (53.92%), and ImageNet-Sketch (41.04%) without extra data, using only the ViT-B/16 backbone and our pyramid adversarial training. Our code is publicly available at pyramidat.github.io.

Unsupervised domain adaptation (UDA) makes predictions for the target domain data while manual annotations are only available in the source domain. Previous methods minimize the domain discrepancy neglecting the class information, which may lead to misalignment and poor generalization performance. To tackle this issue, this paper proposes contrastive adaptation network (CAN) that optimizes a new metric named Contrastive Domain Discrepancy explicitly modeling the intra-class domain discrepancy and the inter-class domain discrepancy. To optimize CAN, two technical issues need to be addressed: 1) the target labels are not available; and 2) the conventional mini-batch sampling is imbalanced. Thus we design an alternating update strategy to optimize both the target label estimations and the feature representations. Moreover, we develop class-aware sampling to enable more efficient and effective training. Our framework can be generally applied to the single-source and multi-source domain adaptation scenarios. In particular, to deal with multiple source domain data, we propose: 1) multi-source clustering ensemble which exploits the complementary knowledge of distinct source domains to make more accurate and robust target label estimations; and 2) boundary-sensitive alignment to make the decision boundary better fitted to the target. Experiments are conducted on three real-world benchmarks (i.e., Office-31 and VisDA-2017 for the single-source scenario, DomainNet for the multi-source scenario). All the results demonstrate that our CAN performs favorably against the state-of-the-art methods. Ablation studies also verify the effectiveness of each key component of our proposed system.

Recent years have witnessed the rapid progress of generative adversarial networks (GANs). However, the success of the GAN models hinges on a large amount of training data. This work proposes a regularization approach for training robust GAN models on limited data. We theoretically show a connection between the regularized loss and an f-divergence called LeCam-divergence, which we find is more robust under limited training data. Extensive experiments on several benchmark datasets demonstrate that the proposed regularization scheme 1) improves the generalization performance and stabilizes the learning dynamics of GAN models under limited training data, and 2) complements the recent data augmentation methods. These properties facilitate training GAN models to achieve state-of-theart performance when only limited training data of the ImageNet benchmark is available. The source code is available at https://github.com/google/lecam-gan.

It has been shown that deep neural networks are prone to overfitting on biased training data. Towards addressing this issue, meta-learning employs a meta model for correcting the training bias. Despite the promising performances, super slow training is currently the bottleneck in the meta learning approaches. In this paper, we introduce a novel Faster Meta Update Strategy (FaMUS) to replace the most expensive step in the meta gradient computation with a faster layer-wise approximation. We empirically find that FaMUS yields not only a reasonably accurate but also a low-variance approximation of the meta gradient. We conduct extensive experiments to verify the proposed method on two tasks. We show our method is able to save two-thirds of the training time while still maintaining the comparable or achieving even better generalization performance. In particular, our method achieves the state-of-the-art performance on both synthetic and realistic noisy labels, and obtains promising performance on long-tailed recognition on standard benchmarks. Code are released at https://github.com/youjiangxu/FaMUS.

Self-supervised pre-training of text representations has been successfully applied to low-resource Neural Machine Translation (NMT). However, it usually fails to achieve notable gains on resource-rich NMT. In this paper, we propose a joint training approach, F2-XEnDec, to combine self-supervised and supervised learning to optimize NMT models. To exploit complementary self-supervised signals for supervised learning, NMT models are trained on examples that are interbred from monolingual and parallel sentences through a new process called crossover encoder-decoder. Experiments on two resource-rich translation benchmarks, WMT’14 English-German and WMT’14 English-French, demonstrate that our approach achieves substantial improvements over several strong baseline methods and obtains a new state of the art of 46.19 BLEU on English-French when incorporating back translation. Results also show that our approach is capable of improving model robustness to input perturbations such as code-switching noise which frequently appears on the social media.

n recent years, text-guided image manipulation has gained increasing attention in the multimedia and computer vision community. The input to conditional image generation has evolved from image-only to multimodality. In this paper, we study a setting that allows users to edit an image with multiple objects using complex text instructions to add, remove, or change the objects. The inputs of the task are multimodal including (1) a reference image and (2) an instruction in natural language that describes desired modifications to the image. We propose a GAN-based method to tackle this problem. The key idea is to treat text as neural operators to locally modify the image feature. We show that the proposed model performs favorably against recent strong baselines on three public datasets. Specifically, it generates images of greater fidelity and semantic relevance, and when used as a image query, leads to better retrieval performance.

Image extrapolation aims at expanding the narrow field of view of a given image patch. Existing models mainly deal with natural scene images of homogeneous regions and have no control of the content generation process. In this work, we study conditional image extrapolation to synthesize new images guided by the input structured text. The text is represented as a graph to specify the objects and their spatial relation to the unknown regions of the image. Inspired by drawing techniques, we propose a progressive generative model of three stages, i.e., generating a coarse bounding-boxes layout, refining it to a finer segmentation layout, and mapping the layout to a realistic output. Such a multi-stage design is shown to facilitate the training process and generate more controllable results. We validate the effectiveness of the proposed method on the face and human clothing dataset in terms of visual results, quantitative evaluations, and flexible controls.

Learning exists in the context of data, yet notions of confidence typically focus on model predictions, not label quality. Confident learning (CL) is an alternative approach which focuses instead on label quality by characterizing and identifying label errors in datasets, based on the principles of pruning noisy data, counting with probabilistic thresholds to estimate noise, and ranking examples to train with confidence. Whereas numerous studies have developed these principles independently, here, we combine them, building on the assumption of a class-conditional noise process to directly estimate the joint distribution between noisy (given) labels and uncorrupted (unknown) labels. This results in a generalized CL which is provably consistent and experimentally performant. We present sufficient conditions where CL exactly finds label errors, and show CL performance exceeding seven recent competitive approaches for learning with noisy labels on the CIFAR dataset. Uniquely, the CL framework is not coupled to a specific data modality or model (e.g., we use CL to find several label errors in the presumed error-free MNIST dataset and improve sentiment classification on text data in Amazon Reviews). We also employ CL on ImageNet to quantify ontological class overlap (e.g., estimating 645 missile images are mislabeled as their parent class projectile), and moderately increase model accuracy (e.g., for ResNet) by cleaning data prior to training. These results are replicable using the open-source cleanlab release.

In this paper, we propose a new adversarial augmentation method for Neural Machine Translation (NMT).The main idea is to minimize the vicinal risk over virtual sentences sampled from two vicinity distributions, of which the crucial one is a novel vicinity distribution for adversarial sentences that describes a smooth interpolated embedding space centered around observed training sentence pairs.We then discuss our approach, AdvAug, to train NMT models using the embeddings of virtual sentences in sequence-tosequence learning.Experiments on Chinese-English, English-French, and English-German translation benchmarks show that AdvAug achieves significant improvements over the Transformer (up to 4.9 BLEU points), and substantially outperforms other data augmentation techniques (e.g.back-translation) without using extra corpora.

This paper studies the problem of predicting the distribution over multiple possible future paths of people as they move through various visual scenes. We make two main contributions. The first contribution is a new dataset, created in a realistic 3D simulator, which is based on real world trajectory data, and then extrapolated by human annotators to achieve different latent goals. This provides the first benchmark for quantitative evaluation of the models to predict multi-future trajectories. The second contribution is a new model to generate multiple plausible future trajectories, which contains novel designs of using multi-scale location encodings and convolutional RNNs over graphs. We refer to our model as Multiverse. We show that our model achieves the best results on our dataset, as well as on the real-world VIRAT/ActEV dataset (which just contains one possible future).

Performing controlled experiments on noisy data is essential in understanding deep learning across noise levels. Due to the lack of suitable datasets, previous research has only examined deep learning on controlled synthetic label noise, and real-world label noise has never been studied in a controlled setting. This paper makes three contributions. First, we establish the first benchmark of controlled real-world label noise from the web. This new benchmark enables us to study the web label noise in a controlled setting for the first time. The second contribution is a simple but effective method to overcome both synthetic and real noisy labels. We show that our method achieves the best result on our dataset as well as on two public benchmarks (CIFAR and WebVision). Third, we conduct the largest study by far into understanding deep neural networks trained on noisy labels across different noise levels, noise types, network architectures, and training settings.

In Embodied Question Answering (EmbodiedQA), an agent interacts with an environment to gather necessary information for answering user questions. Existing works have laid a solid foundation towards solving this interesting problem. But the current performance, especially in navigation, suggests that EmbodiedQA might be too challenging for the contemporary approaches. In this paper, we empirically study this problem and introduce 1) a simple yet effective baseline that achieves promising performance; 2) an easier and practical setting for EmbodiedQA where an agent has a chance to adapt the trained model to a new environment before it actually answers users questions. In this new setting, we randomly place a few objects in new environments, and upgrade the agent policy by a distillation network to retain the generalization ability from the trained model. On the EmbodiedQA v1 benchmark, under the standard setting, our simple baseline achieves very competitive results to the-state-of-the-art; in the new setting, we found the introduced small change in settings yields a notable gain in navigation.

Neural machine translation (NMT) often suffers from the vulnerability to noisy perturbations in the input.We propose an approach to improving the robustness of NMT models, which consists of two parts: (1) attack the translation model with adversarial source examples; (2) defend the translation model with adversarial target inputs to improve its robustness against the adversarial source inputs.For the generation of adversarial inputs, we propose a gradient-based method to craft adversarial examples informed by the translation loss over the clean inputs.Experimental results on Chinese-English and English-German translation tasks demonstrate that our approach achieves significant improvements (2.8 and 1.6 BLEU points) over Transformer on standard clean benchmarks as well as exhibiting higher robustness on noisy data.

Unsupervised Domain Adaptation (UDA) makes predictions for the target domain data while manual annotations are only available in the source domain. Previous methods minimize the domain discrepancy neglecting the class information, which may lead to misalignment and poor generalization performance. To address this issue, this paper proposes Contrastive Adaptation Network (CAN) optimizing a new metric which explicitly models the intra-class domain discrepancy and the inter-class domain discrepancy. We design an alternating update strategy for training CAN in an end-to-end manner. Experiments on two real-world benchmarks Office-31 and VisDA-2017 demonstrate that CAN performs favorably against the state-of-the-art methods and produces more discriminative features.

Deciphering human behaviors to predict their future paths/trajectories and what they would do from videos is important in many applications. Motivated by this idea, this paper studies predicting a pedestrian's future path jointly with future activities. We propose an end-to-end, multi-task learning system utilizing rich visual features about human behavioral information and interaction with their surroundings. To facilitate the training, the network is learned with an auxiliary task of predicting future location in which the activity will happen. Experimental results demonstrate our state-of-the-art performance over two public benchmarks on future trajectory prediction. Moreover, our method is able to produce meaningful future activity prediction in addition to the path. The result provides the first empirical evidence that joint modeling of paths and activities benefits future path prediction.

In this paper, we study the task of image retrieval, where the input query is specified in the form of an image plus some text that describes desired modifications to the input image. For example, we may present an image of the Eiffel tower, and ask the system to find images which are visually similar, but are modified in small ways, such as being taken at nighttime instead of during the day. o tackle this task, we embed the query (reference image plus modification text) and the target (images). The encoding function of the image text query learns a representation, such that the similarity with the target image representation is high iff it is a ``positive match''. We propose a new way to combine image and text through residual connection, that is designed for this retrieval task. We show this outperforms existing approaches on 3 different datasets, namely Fashion-200k, MIT-States and a new synthetic dataset we create based on CLEVR. We also show that our approach can be used to perform image classification with compositionally novel labels, and we outperform previous methods on MIT-States on this task.

Recent insights on language and vision with neural networks have been successfully applied to simple single-image visual question answering. However, to tackle real-life question answering problems on multimedia collections such as personal photo albums, we have to look at whole collections with sequences of photos. This paper proposes a new multimodal MemexQA task: given a sequence of photos from a user, the goal is to automatically answer questions that help users recover their memory about an event captured in these photos. In addition to a text answer, a few grounding photos are also given to justify the answer. The grounding photos are necessary as they help users quickly verifying the answer. Towards solving the task, we 1) present the MemexQA dataset, the first publicly available multimodal question answering dataset consisting of real personal photo albums; 2) propose an end-to-end trainable network that makes use of a hierarchical process to dynamically determine what media and what time to focus on in the sequential data to answer the question. Experimental results on the MemexQA dataset demonstrate that our model outperforms strong baselines and yields the most relevant grounding photos on this challenging task.

Recent insights on language and vision with neural networks have been successfully applied to simple single-image visual question answering. However, to tackle real-life question answering problems on multimedia collections such as personal photos, we have to look at whole collections with sequences of photos or videos. When answering questions from a large collection, a natural problem is to identify snippets to support the answer. In this paper, we describe a novel neural network called Focal Visual-Text Attention network (FVTA) for collective reasoning in visual question answering, where both visual and text sequence information such as images and text metadata are presented. FVTA introduces an end-to-end approach that makes use of a hierarchical process to dynamically determine what media and what time to focus on in the sequential data to answer the question. FVTA can not only answer the questions well but also provides the justifications which the system results are based upon to get the answers. FVTA achieves state-of-the-art performance on the MemexQA dataset and competitive results on the MovieQA dataset.

Recent deep networks are capable of memorizing the entire data even when the labels are completely random. To overcome the overfitting on corrupted labels, we propose a novel technique of learning another neural network, called MentorNet, to supervise the training of the base deep networks, namely, StudentNet. During training, MentorNet provides a curriculum (sample weighting scheme) for StudentNet to focus on the sample the label of which is probably correct. Unlike the existing curriculum that is usually predefined by human experts, MentorNet learns a data-driven curriculum dynamically with StudentNet. Experimental results demonstrate that our approach can significantly improve the generalization performance of deep networks trained on corrupted training data. Notably, to the best of our knowledge, we achieve the best-published result on WebVision, a large benchmark containing 2.2 million images of real-world noisy labels.

Image captioning is a challenging task where the machine automatically describes an image by sentences or phrases. It often requires a large number of paired image-sentence annotations for training. However, a pre-trained captioning model can hardly be applied to a new domain in which some novel object categories exist, i.e., the objects and their description words are unseen during model training. To correctly caption the novel object, it requires professional human workers to annotate the images by sentences with the novel words. It is labor expensive and thus limits its usage in real-world applications. In this paper, we introduce the zero-shot novel object captioning task where the machine generates descriptions without extra training sentences about the novel object. To tackle the challenging problem, we propose a Decoupled Novel Object Captioner (DNOC) framework that can fully decouple the language sequence model from the object descriptions. DNOC has two components. 1) A Sequence Model with the Placeholder (SM-P) generates a sentence containing placeholders. The placeholder represents an unseen novel object. Thus, the sequence model can be decoupled from the novel object descriptions. 2) A key-value object memory built upon the freely available detection model, contains the visual information and the corresponding word for each object. A query generated from the SM-P is used to retrieve the words from the object memory. The placeholder will further be filled with the correct word, resulting in a caption with novel object descriptions. The experimental results on the held-out MSCOCO dataset demonstrate the ability of DNOC in describing novel concepts.

The boom of mobile devices and cloud services has led to an explosion of personal photo and video data. However, due to the missing user-generated metadata such as titles or descriptions, it usually takes a user a lot of swipes to find some video on the cell phone. To solve the problem, we present an innovative idea called Visual Memory QA which allow a user not only to search but also to ask questions about her daily life captured in the personal videos. The proposed system automatically analyzes the content of personal videos without user-generated metadata, and offers a conversational interface to accept and answer questions. To the best of our knowledge, it is the first to answer personal questions discovered in personal photos or videos. The example questions are "what was the lat time we went hiking in the forest near San Francisco?"; "did we have pizza last week?"; "with whom did I have dinner in AAAI 2015?".

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.

Given any complicated or specialized video content search query, e.g. ”Batkid (a kid in batman costume)” or ”destroyed buildings”, existing methods require manually labeled data to build detectors for searching. We present a demonstration of an artificial intelligence application, Webly-labeled Learning (WELL) that enables learning of ad-hoc concept detectors over unlimited Internet videos without any manual an-notations. A considerable number of videos on the web are associated with rich but noisy contextual information, such as the title, which provides a type of weak annotations or la-bels of the video content. To leverage this information, our system employs state-of-the-art webly-supervised learning(WELL) (Liang et al. ). WELL considers multi-modal information including deep learning visual, audio and speech features, to automatically learn accurate video detectors based on the user query. The learned detectors from a large number of web videos allow users to search relevant videos over their personal video archives, not requiring any textual metadata,but as convenient as searching on Youtube.

With the explosion in the availability of user-generated videos documenting any conflicts and human rights abuses around the world, analysts and researchers increasingly find themselves overwhelmed with massive amounts of video data to acquire and analyze useful information. In this paper, we develop a temporal localization framework for intense audio events in videos which addresses the problem. The proposed method utilizes Localized Self-Paced Reranking (LSPaR) to refine the localization results. LSPaR utilizes samples from easy to noisier ones so that it can overcome the noisiness of the initial retrieval results from user-generated videos. We show our framework's efficacy on localizing intense audio event like gunshot, and further experiments also indicate that our methods can be generalized to localizing other audio events in noisy videos.

Learning video concept detectors automatically from the big but noisy web data with no additional manual annotations is a novel but challenging area in the multimedia and the machine learning community. A considerable amount of videos on the web is associated with rich but noisy contextual information, such as the title and other multi-modal information, which provides weak annotations or labels about the video content. To tackle the problem of large-scale noisy learning, We propose a novel method called Multi-modal WEbly-Labeled Learning (WELL-MM), which is established on the state-of-the-art machine learning algorithm inspired by the learning process of human. WELL-MM introduces a novel multi-modal approach to incorporate meaningful prior knowledge called curriculum from the noisy web videos. We empirically study the curriculum constructed from the multi-modal features of the Internet videos and images. The comprehensive experimental results on FCVID and YFCC100M demonstrate that WELL-MM outperforms state-of-the-art studies by a statically significant margin on learning concepts from noisy web video data. In addition, the results also verify that WELL-MM is robust to the level of noisiness in the video data. Notably, WELL-MM trained on sufficient noisy web labels is able to achieve a better accuracy to supervised learning methods trained on the clean manually labeled data.

The ubiquity of mobile devices and cloud services has led to an unprecedented growth of online personal photo and video collections. Due to the scarcity of personal media search log data, research to date has mainly focused on searching images and videos on the web. However, in order to manage the exploding amount of personal photos and videos, we raise a fundamental question: what are the differences and similarities when users search their own photos versus the photos on the web? To the best of our knowledge, this paper is the first to study personal media search using large-scale real-world search logs. We analyze different types of search sessions mined from Flickr search logs and discover a number of interesting characteristics of personal media search in terms of information needs and click behaviors. The insightful observations will not only be instrumental in guiding future personal media search methods, but also benefit related tasks such as personal photo browsing and recommendation. Our findings suggest there is a significant gap between personal queries and automatically detected concepts, which is responsible for the low accuracy of many personal media search queries. To bridge the gap, we propose the deep query understanding model to learn a mapping from the personal queries to the concepts in the clicked photos. Experimental results verify the efficacy of the proposed method in improving personal media search, where the proposed method consistently outperforms baseline methods.

Recent progresses in multimedia event detection have enabled us to find videos about a predefined event from a large-scale video collection. Research towards more intrinsic unsupervised video understanding is an interesting but understudied field. Specifically, given a collection of videos sharing a common event of interest, the goal is to discover the salient fragments, i.e., the curt video fragments that can concisely portray the underlying event of interest, from each video. To explore this novel direction, this paper proposes an unsupervised event saliency revealing framework. It first extracts features from multiple modalities to represent each shot in the given video collection. Then, these shots are clustered to build the cluster-level event saliency revealing framework, which explores useful information cues (i.e., the intra-cluster prior, inter-cluster discriminability, and inter-cluster smoothness) by a concise optimization model. Compared with the existing methods, our approach could highlight the intrinsic stimulus of the unseen event within a video in an unsupervised fashion. Thus, it could potentially benefit to a wide range of multimedia tasks like video browsing, understanding, and search. To quantitatively verify the proposed method, we systematically compare the method to a number of baseline methods on the TRECVID benchmarks. Experimental results have demonstrated its effectiveness and efficiency.

The Internet has been witnessing an explosion of video content. According to a Cisco study, video content is estimated to account for 80% of all the entire world's internet traffic by 2019. Video data are becoming one of the most valuable sources to assess information and knowledge. However, existing video search solutions are still based on text matching (text-to-text search), and could fail for the huge volumes of videos that have little relevant metadata or no metadata at all. The need for large-scale and intelligent video search, which bridges the gap between the user's information need and the video content, seems to be urgent.

Curriculum learning (CL) or self-paced learning (SPL) represents a recently proposed learning regime inspired by the learning process of humans and animals that gradually proceeds from easy to more complex samples in training. The two methods share a similar conceptual learning paradigm, but differ in specific learning schemes. In CL, the curriculum is predetermined by prior knowledge, and remain fixed thereafter. Therefore, this type of method heavily relies on the quality of prior knowledge while ignoring feedback about the learner. In SPL, the curriculum is dynamically determined to adjust to the learning pace of the leaner. However, SPL is unable to deal with prior knowledge, rendering it prone to overfitting. In this paper, we discover the missing link between CL and SPL, and propose a unified framework named self-paced curriculum leaning (SPCL). SPCL is formulated as a concise optimization problem that takes into account both prior knowledge known before training and the learning progress during training. In comparison to human education, SPCL is analogous to "instructor-student-collaborative" learning mode, as opposed to "instructor-driven" in CL or "student-driven" in SPL. Empirically, we show that the advantage of SPCL on two tasks.

Matrix factorization (MF) has been attracting much attention due to its wide applications. However, since MF models are generally non-convex, most of the existing methods are easily stuck into bad local minima, especially in the presence of outliers and missing data. To alleviate this deficiency, in this study we present a new MF learning methodology by gradually including matrix elements into MF training from easy to complex. This corresponds to a recently proposed learning fashion called self-paced learning (SPL), which has been demonstrated to be beneficial in avoiding bad local minima. We also generalize the conventional binary (hard) weighting scheme for SPL to a more effective real-valued (soft) weighting manner. The effectiveness of the proposed self-paced MF method is substantiated by a series of experiments on synthetic, structure from motion and background subtraction data.

As an interesting and emerging topic, co-saliency detection aims at simultaneously extracting common salient objects in a group of images. Traditional co-saliency detection approaches rely heavily on human knowledge for designing hand-crafted metrics to explore the intrinsic patterns underlying co-salient objects. Such strategies, however, always suffer from poor generalization capability to flexibly adapt various scenarios in real applications, especially due to their lack of insightful understanding of the biological mechanisms of human visual co-attention. To alleviate this problem, we propose a novel framework for this task, by naturally reformulating it as a multiple-instance learning (MIL) problem and further integrating it into a self-paced learning (SPL) regime. The proposed framework on one hand is capable of fitting insightful metric measurements and discovering common patterns under co-salient regions in a self-learning way by MIL, and on the other hand tends to promise the learning reliability and stability by simulating the human learning process through SPL. Experiments on benchmark datasets have demonstrated the effectiveness of the proposed framework as compared with the state-of-the-arts.

Semantic search in video is a novel and challenging problem in information and multimedia retrieval. Existing solutions are mainly limited to text matching, in which the query words are matched against the textual metadata generated by users. This paper presents a state-of-the-art system for event search without any textual metadata or example videos. The system relies on substantial video content understanding and allows for semantic search over a large collection of videos. The novelty and practicality is demonstrated by the evaluation in NIST TRECVID 2014, where the proposed system achieves the best performance. We share our observations and lessons in building such a state-of-the-art system, which may be instrumental in guiding the design of the future system for semantic search in video.

Recent improvements in content-based video search have led to systems with promising accuracy, thus opening up the possibility for interactive content-based video search to the general public. We present an interactive system based on a state-of-the-art content-based video search pipeline which enables users to do multimodal text-to-video and video-to-video search in large video collections, and to incrementally refine queries through relevance feedback and model visualization. Also, the comprehensive functionalities enhance a flexible formulation of multimodal queries with different characteristics. Quantitative and qualitative analysis shows that our system is capable of assisting users to incrementally build effective queries over complex event topics.

Many content-based video search (CBVS) systems have been proposed to analyze the rapidly-increasing amount of user-generated videos on the Internet. Though the accuracy of CBVS systems have drastically improved, these high accuracy systems tend to be too inefficient for interactive search. Therefore, to strive for real-time web-scale CBVS, we perform a comprehensive study on the different components in a CBVS system to understand the trade-offs between accuracy and speed of each component. Directions investigated include exploring different low-level and semantics-based features, testing different compression factors and approximations during video search, and understanding the time v.s. accuracy trade-off of reranking. Extensive experiments on data sets consisting of more than 1,000 hours of video showed that through a combination of effective features, highly compressed representations, and one iteration of reranking, our proposed system can achieve an 10,000-fold speedup while retaining 80% accuracy of a state-of-the-art CBVS system. We further performed search over 1 million videos and demonstrated that our system can complete the search in 0.975 seconds with a single core, which potentially opens the door to interactive web-scale CBVS for the general public.

Large-scale content-based semantic search in video is an interesting and fundamental problem in multimedia analysis and retrieval. Existing methods index a video by the raw concept detection score that is dense and inconsistent, and thus cannot scale to "big data" that are readily available on the Internet. This paper proposes a scalable solution. The key is a novel step called concept adjustment that represents a video by a few salient and consistent concepts that can be efficiently indexed by the modified inverted index. The proposed adjustment model relies on a concise optimization framework with interpretations. The proposed index leverages the text-based inverted index for video retrieval. Experimental results validate the efficacy and the efficiency of the proposed method. The results show that our method can scale up the semantic search while maintaining state-of-the-art search performance. Specifically, the proposed method (with reranking) achieves the best result on the challenging TRECVID Multimedia Event Detection (MED) zero-example task. It only takes 0.2 second on a single CPU core to search a collection of 100 million Internet videos.

We propose a novel method MultiModal Pseudo Relevance Feedback (MMPRF) for event search in video, which requires no search examples from the user. Pseudo Relevance Feedback has shown great potential in retrieval tasks, but previous works are limited to unimodal tasks with only a single ranked list. To tackle the event search task which is inherently multimodal, our proposed MMPRF takes advantage of multiple modalities and multiple ranked lists to enhance event search performance in a principled way. The approach is unique in that it leverages not only semantic features, but also non-semantic low-level features for event search in the absence of training data. Evaluated on the TRECVID MEDTest dataset, the approach improves the baseline by up to 158% in terms of the mean average precision. It also significantly contributes to CMU Team's final submission in TRECVID-13 Multimedia Event Detection.

Viral videos that gain popularity through the process of Internet sharing are having a profound impact on society. Existing studies on viral videos have only been on small or confidential datasets. We collect by far the largest open benchmark for viral video study called CMU Viral Video Dataset, and share it with researchers from both academia and industry. Having verified existing observations on the dataset, we discover some interesting characteristics of viral videos. Based on our analysis, in the second half of the paper, we propose a model to forecast the future peak day of viral videos. The application of our work is not only important for advertising agencies to plan advertising campaigns and estimate costs, but also for companies to be able to quickly respond to rivals in viral marketing campaigns. The proposed method is unique in that it is the first attempt to incorporate video metadata into the peak day prediction. The empirical results demonstrate that the proposed method outperforms the state-of-the-art methods, with statistically significant differences.

Spatial Pyramid Matching (SPM) assumes that the spatial Bag-of-Words (BoW) representation is independent of data. However, evidence has shown that the assumption usually leads to a suboptimal representation. In this paper, we propose a novel method called Jensen-Shannon (JS) Tiling to learn the BoW representation from data directly at the BoW level. The proposed JS Tiling is especially appropriate for large-scale datasets as it is orders of magnitude faster than existing methods, but with comparable or even better classification precision. Experimental results on four benchmarks including two TRECVID12 datasets validate that JS Tiling outperforms the SPM and the state-of-the-art methods. The runtime comparison demonstrates that selecting BoW representations by JS Tiling is more than 1,000 times faster than running classifiers. Besides, JS Tiling is an important component contributing to CMU Teams' final submission in TRECVID 2012 Multimedia Event Detection.

Reranking has been a focal technique in multimedia retrieval due to its efficacy in improving initial retrieval results. Current reranking methods, however, mainly rely on the heuristic weighting. In this paper, we propose a novel reranking approach called Self-Paced Reranking (SPaR) for multimodal data. As its name suggests, SPaR utilizes samples from easy to more complex ones in a self-paced fashion. SPaR is special in that it has a concise mathematical objective to optimize and useful properties that can be theoretically verified. It on one hand offers a unified framework providing theoretical justifications for current reranking methods, and on the other hand generates a spectrum of new reranking schemes. This paper also advances the state-of-the-art self-paced learning research which potentially benefits applications in other fields. Experimental results validate the efficacy and the efficiency of the proposed method on both image and video search tasks. Notably, SPaR achieves by far the best result on the challenging TRECVID multimedia event search task.

Online instructional videos have become a popular way for people to learn new skills encompassing art, cooking and sports. As watching instructional videos is a natural way for humans to learn, analogously, machines can also gain knowledge from these videos. We propose to utilize the large amount of instructional videos available online to harvest examples of various actions in an unsupervised fashion. The key observation is that in instructional videos, the instructor's action is highly correlated with the instructor's narration. By leveraging this correlation, we can exploit the timing of action corresponding terms in the speech transcript to temporally localize actions in the video and harvest action examples. The proposed method is scalable as it requires no human intervention. Experiments show that the examples harvested are of reasonably good quality, and action detectors trained on data collected by our unsupervised method yields comparable performance with detectors trained with manually collected data on the TRECVID Multimedia Event Detection task.

Speaker adaptive training (SAT) is a well studied technique for Gaussian mixture acoustic models (GMMs). Recently we proposed to perform SAT for deep neural networks (DNNs), with speaker i-vectors applied in feature learning. The resulting SAT-DNN models significantly outperform DNNs on word error rates (WERs). In this paper, we present different methods to further improve and extend SAT-DNN. First, we conduct detailed analysis to investigate i-vector extractor training and flexible feature fusion. Second, the SAT-DNN approach is extended to improve tasks including bottleneck feature (BNF) generation, convolutional neural network (CNN) acoustic modeling and multilingual DNN-based feature extraction. Third, for transcribing multimedia data, we enrich the i-vector representation with global speaker attributes (age, gender, etc.) obtained automatically from video signals. On a collection of instructional videos, incorporation of the additional visual features is observed to boost the recognition accuracy of SAT-DNN.

This paper addresses the challenge of Multimedia Event Detection by proposing a novel method for high-level and low-level features fusion based on collective classification. Generally, the method consists of three steps: training a classifier from low-level features; encoding high-level features into graphs; and diffusing the scores on the established graph to obtain the final prediction. The final prediction is derived from multiple graphs each of which corresponds to a high-level feature. The paper investigates two graph construction methods using logarithmic and exponential loss functions, respectively and two collective classification algorithms, i.e. Gibbs sampling and Markov random walk. The theoretical analysis demonstrates that the proposed method converges and is computationally scalable and the empirical analysis on TRECVID 2011 Multimedia Event Detection dataset validates its outstanding performance compared to state-of-the-art methods, with an added benefit of interpretability.