Sequential recommendation aims to predict the next item based on historical interactions. To further enhance the reasoning capability in sequential recommendation, LLMs are employed to predict the next item or generate semantic IDs for item representation, given LLMs' extensive domain knowledge and reasoning ability. However, existing LLM-based methods suffer from two limitations. (i) The scarcity of recommendation data with reasoning paths makes it challenging to design suitable chain-of-thought prompting templates, and the full potential of LLMs' reasoning abilities remains underutilized. (ii) Upon obtaining semantic IDs, the LLMs and their representations are excluded from the subsequent recommendation model training, preventing downstream models from fully utilizing the rich semantic information encoded within these IDs. To address these issues, we propose a novel CoderRec framework, which is capable of fully exploiting the information encoded in semantic IDs to guide the recommendation process. Specifically, to address the problem of scarcity in reasoning path-augmented data, we introduce latent reasoning into sequential recommendation and treat the representation captured by the downstream model as domain-specific latent thought, enabling implicit logical inference without requiring explicit CoT annotations. To ensure that the downstream recommendation models are able to deeply leverage the semantic information within IDs, we propose a novel cross-scale model collaboration strategy, which employs cross-scale IDs and a two-phase approach to align LLM-derived semantics with recommendation objectives. Extensive experiments have shown the effectiveness of our proposed CoderRec framework.

Video Large Language Models (video LLMs) have demonstrated remarkable capabilities in video understanding tasks, such as video question answering and temporal localization. However, understanding long videos still remains a significant challenge. Existing video LLMs adopt uni-granularity tokens for long videos, failing to simultaneously understand both high-level semantics and low-level visual details in videos. To tackle this problem, we propose ReCrossVLLM, a reflective cross-granularity grounding framework for video LLM with preference optimization to collaboratively achieve long video understanding, which not only retains the capabilities of high-level video semantics understanding, but also strengthens the fine-grained understanding abilities. Specifically, we propose the coarse-to-fine grounding and fine-to-coarse reflection strategies for long video understanding. In the coarse-to-fine grounding strategy, the video LLM with a coarse-grained module first locates the key video segments from the long video by tackling massive frames of the long video with fewer per-frame tokens. And then video LLM adapted with the fine-grained module further analyzes the key video segments with more per-frame tokens so that it can understand fine-grained information. In case the video LLM locates the wrong key video segments, during the inference stage, our designed fine-to-coarse reflection strategy instructs the fine-grained module to reflect the effectiveness of the locating result and decide whether to return to the coarse-to-fine grounding strategy with reflection feedback. Additionally, during the training stage, the coarse-to-fine grounding strategy is optimized with our proposed cross-granularity preference optimization strategy to further improve grounding efficiency. Extensive experiments for long video question answering and temporal video grounding tasks demonstrate that our proposed ReCrossVLLM framework can significantly improve the Video Large Language Model for long video understanding.

Large Vision-Language Models have drawn much attention and become increasingly applicable in complicated multimodal tasks such as visual question answering, video grounding, etc. However, it still suffers from inefficiency problem during the inference stage due to the computational overhead brought by the large number of visual tokens. Existing works either utilize an attention score (or visual-text relevance) to filter out the less significant visual tokens, or insert learnable projection layers to directly compress the tokens, which neglects the informative details in visual signals and introduces information loss, resulting in poor generalizability to test data. To solve these problems, in this paper we propose a novel Disentangled Visual Token Compression module, i.e., DiViCo, that effectively compresses the visual tokens and maintains good performance simultaneously. In concrete, we first select the top τ% visual tokens according to their average attention scores, then predict the gap between these selected tokens and the original information by employing the chosen tokens in a disentangled and variational manner. Specifically, we model the mean and variance, sampling the predicted gap from the Gaussian prior. We further keep the informativeness of the compressed visual tokens via KL divergence, which ensures the generalizability of the model. Extensive experiments demonstrate the advantage of our proposed DiViCo module against several state-of-the-art baselines over various real-world datasets. Most notably, LLaVA-v1.5-7b equipped with DiViCo is able to reduce 67.7% FLOPs and save 51.7% time while maintaining 95.6% of the accuracy for LLaVA-v1.5-7b without any compression.

Multi-modal generative AI (Artificial Intelligence) has attracted increasing attention from both academia and industry. Particularly, two dominant families of techniques have emerged: i) Multi-modal large language models (LLMs) demonstrate impressive ability formulti-modal understanding; and ii) Diffusion models exhibit remarkable multi-modal powers in terms ofmulti-modal generation. Therefore, this paper provides a comprehensive overview of multi-modal generative AI, including multi-modal LLMs, diffusions, and the unification for understanding and generation. To lay a solid foundation for unified models, we first provide a detailed review of both multi-modal LLMs and diffusion models, respectively, including their probabilistic modeling procedure, multi-modal architecture design, and advanced applications to image/video LLMs as well as text-to-image/video generation. Furthermore, we explore the emerging efforts toward unified models for understanding and generation. To achieve the unification of understanding and generation, we investigate key designs including autoregressive-based and diffusion-based modeling, as well as dense and Mixture-of-Experts (MoE) architectures. We then introduce several strategies for unified models, analyzing their potential advantages and disadvantages. In addition, we summarize the common datasets widely used for multi-modal generative AI pretraining. Last but not least, we present several challenging future research directions that may contribute to the ongoing advancement of multi-modal generative AI.

Cross-domain recommendation (CDR) mitigates data sparsity and cold-start issues in recommendation systems. While recent CDR approaches using graph neural networks (GNNs) capture complex user-item interactions, they rely on manually designed architectures that are often suboptimal and labor-intensive. Additionally, extracting valuable behavioral information from source domains to improve target domain recommendations remains challenging. To address these challenges, we propose Behavior importance-aware Graph Neural Architecture Search (BiGNAS), a framework that jointly optimizes GNN architecture and data importance for CDR. BiGNAS introduces two key components: a Cross-Domain Customized Supernetwork and a Graph-Based Behavior Importance Perceptron. The supernetwork, as a one-shot, retrain-free module, automatically searches the optimal GNN architecture for each domain without the need for retraining. The perceptron uses auxiliary learning to dynamically assess the importance of source domain behaviors, thereby improving target domain recommendations. Extensive experiments on benchmark CDR datasets and a large-scale industry advertising dataset demonstrate that BiGNAS consistently outperforms state-of-the-art baselines. To the best of our knowledge, this is the first work to jointly optimize GNN architecture and behavior data importance for cross-domain recommendation.

Video corpus grounding (VCG), which aims to retrieve relevant video moments from a video corpus, has attracted significant attention in the multimedia research community. However, the existing VCG setting primarily focuses on matching textual descriptions with videos and ignores the distinct visual identities in the videos, thus resulting in inaccurate understanding of video content and deteriorated retrieval performances. To address this limitation, we introduce a novel task, Identity-Text Video Corpus Grounding (ITVCG), which simultaneously utilize textual descriptions and visual identities as queries. As such, ITVCG benefits in enabling more accurate video corpus grounding with visual identities, as well as providing users with more flexible options to locate relevant frames based on either textual descriptions or textual descriptions and visual identities. To conduct evaluations regarding the novel ITVCG task, we propose the TVR-IT dataset, comprising 463 identity images from 6 TV shows, with 68,840 out of 72,840 queries containing at least one identity image. Furthermore, we propose Video-Locator, the first model designed for the ITVCG task. Our proposed Video-Locator integrates video-identity-text alignment and multi-modal fine-grained fusion components, facilitating a video large language model (Video LLM) to jointly understand textual descriptions, visual identities, as well as videos. Experimental results demonstrate the effectiveness of the proposed Video-Locator model and highlight the importance of identity-generalization capability for ITVCG.

Text-to-Video generation, which utilizes the provided text prompt to generate high-quality videos, has drawn increasing attention and achieved great success due to the development of diffusion models recently. Existing methods mainly rely on a pre-trained text encoder to capture the semantic information and perform cross attention with the encoded text prompt to guide the generation of video. However, when it comes to complex prompts that contain dynamic scenes and multiple camera-view transformations, these methods can not decompose the overall information into separate scenes, as well as fail to smoothly change scenes based on the corresponding camera-views. To solve these problems, we propose a novel method, i.e., Modular-Cam. Specifically, to better understand a given complex prompt, we utilize a large language model to analyze user instructions and decouple them into multiple scenes together with transition actions. To generate a video containing dynamic scenes that match the given camera-views, we incorporate the widely-used temporal transformer into the diffusion model to ensure continuity within a single scene and propose CamOperator, a modular network based module that well controls the camera movements. Moreover, we propose AdaControlNet, which utilizes ControlNet to ensure consistency across scenes and adaptively adjusts the color tone of the generated video. Extensive qualitative and quantitative experiments prove our proposed Modular-Cam's strong capability of generating multi-scene videos together with its ability to achieve fine-grained control of camera movements. Generated results are available at https://modular-cam.github.io.

Video Temporal Grounding (VTG) localizes moments in untrimmed videos using natural language queries. Most VTG datasets focus on short videos, and existing approaches excel in short-term cross-modal matching but struggle with long VTG, where long-range temporal reasoning is required for complex events. Existing approaches typically output timestamp predictions without intermediate steps, limiting effective reasoning, whereas humans solve this step by step. To address this, we propose a long VTG framework, StepVTG, with multimodal visual and speech inputs, leveraging Large Language Models (LLMs) for step-by-step reasoning. Specifically, we transform task descriptions, speech, and visual inputs into text prompts. To enhance temporal reasoning, we introduce the Boundary-Perceptive Prompting strategy, which includes: i) a multiscale denoising Chain-of-Thought (CoT) combining global and local semantics with noise filtering, ii) validity principles to ensure LLMs generate reasonable, parsable predictions, and iii) one-shot In-Context Learning (ICL) to improve reasoning via imitation. For evaluation, we establish MM-LVTG, a new long VTG benchmark with multimodal inputs, and demonstrate through extensive experiments that StepVTG achieves state-of-the-art performance. It offers explainable reasoning steps for predictions and reveals potential in facilitating video understanding with off-the-shelf LLMs.

Video generation has witnessed great success recently, but its application in generating long videos still remains challenging due to the difficulty in maintaining the temporal consistency of generated videos and the high memory cost during generation. To tackle the problems, in this paper, we propose a brave and new idea of Multi-sentence Video Grounding for Long Video Generation, connecting the massive video moment retrieval to the video generation task for the first time, providing a new paradigm for long video generation. The method of our work can be summarized as three steps: (i) We design sequential scene text prompts as the queries for video grounding, utilizing the massive video moment retrieval to search for video moment segments that meet the text requirements in the video database. (ii) Based on the source frames of retrieved video moment segments, we adopt video editing methods to create new video content while preserving the temporal consistency of the retrieved video. Since editing can be conducted segment by segment and even frame by frame, it largely reduces memory costs. (iii) We also attempt video morphing and personalized generation methods to improve the subject consistency of long video generation, providing ablation experimental results for the subtasks of long video generation. Our approach seamlessly extends the development in image/video editing, video morphing and personalized generation, and video grounding to the long video generation, offering effective solutions for generating long videos at a low memory cost.

Multimodal Large Language Models (Multimodal LLMs) have shown their strength in Video Question Answering (VideoQA). However, due to the black-box nature of end-to-end training strategies, existing approaches based on Multimodal LLMs suffer from the lack of interpretability for VideoQA: they can neither present reasoning paths nor indicate where the answers are derived from the video. To address this issue, we propose **MSR-ViR** (**M**odularized **S**elf-**R**eflected **Vi**deo **R**easoner), which for the first time integrates modular networks to Multimodal LLMs, capable of providing VideoQA with explicit reasoning paths for more interpretability. Specifically, a **MoST-Grounding** (Modularized Spatial-Temporal Grounding) network is proposed to decompose complex questions via tree-structured policies, localizing relevant temporal and spatial segments within videos through step-by-step reasoning. The proposed MoST-Grounding network provides explicit visually grounded information for Multimodal LLMs with clear reasoning paths, thus enhancing interpretability for the predicted answers. To further improve the reasoning quality, we design an **Alternate Self-reflection Training Strategy** to jointly optimize policy generation and Multimodal LLMs. Experiments on real-world datasets demonstrate the superiority of our proposed MSR-ViR framework in video understanding, reasoning transparency, and providing explicit localization evidence for answers.

Continual multimodal instruction tuning is crucial for adapting Multimodal Large Language Models (MLLMs) to evolving tasks. However, most existing methods adopt a fixed architecture, struggling with adapting to new tasks due to static model capacity. We propose to evolve the architecture under parameter budgets for dynamic task adaptation, which remains unexplored and imposes two challenges: 1) task architecture conflict, where different tasks require varying layer-wise adaptations, and 2) modality imbalance, where different tasks rely unevenly on modalities, leading to unbalanced updates. To address these challenges, we propose a novel Dynamic Mixture of Curriculum LoRA Experts (D-MoLE) method, which automatically evolves MLLM's architecture with controlled parameter budgets to continually adapt to new tasks while retaining previously learned knowledge. Specifically, we propose a dynamic layer-wise expert allocator, which automatically allocates LoRA experts across layers to resolve architecture conflicts, and routes instructions layer-wisely to facilitate knowledge sharing among experts. Then, we propose a gradient-based inter-modal continual curriculum, which adjusts the update ratio of each module in MLLM based on the difficulty of each modality within the task to alleviate the modality imbalance problem. Extensive experiments show that D-MoLE significantly outperforms state-of-the-art baselines, achieving a 15 percent average improvement over the best baseline. To the best of our knowledge, this is the first study of continual learning for MLLMs from an architectural perspective.

Large language models (LLMs) have significantly influenced advancements in sequential recommendation. Nevertheless, the integration and alignment of LLMs with sequence recommenders is often underexploited in current research. Existing LLM-based sequential recommenders mostly rely on textual descriptions, neglecting user visual preferences and suffering from LLM hallucination, which can result in suboptimal recommendations. To address these challenges, we propose AlmostRec, a novel framework that incorporates multimodal information, including historical interaction IDs, textual descriptions, and images of items into large foundation models, to facilitate controllable predictions. Instead of employing a textual LLM, AlmostRec utilizes a large multimodal model (LMM) as a backbone, complemented by a content-behavior guidance module to align multimodal information. The framework's ID prediction objective, enhanced via the parameter-efficient LoRA approach, ensures a principal alignment with the sequential recommendation and is not swayed by hallucinations. AlmostRec effectively bridges the gap between large vision-language models with sequential recommenders, offering contextually relevant predictions in multimodal scenarios. Experimental results on real-world datasets demonstrate the superior performance of AlmostRec compared to both traditional and recent LLM-based recommendation approaches.

Multi-conditional image generation aims to create customized images that align with multiple specified conditions. Existing methods, whether through end-to-end training or by fine-tuning adapters to integrate pre-trained control modules of the same category (e.g., LoRA, IP-Adapter, ControlNet, T2I-Adapter), are restricted to a closed set of predefined input conditions. To overcome this limitation, we propose ModuleTeam, a training-free method for latent mixture of arbitrary control modules, capable of handling open-set conditions by incorporating the corresponding modules. The design of ModuleTeam is rooted in two key findings: (i) modules interfere with each other at the level of model parameters, and (ii) module weights contribute to the generated images by affecting the noise predictions within the diffusion process in an approximately linear manner. The first finding motivates our latent mixture approach, which mixes the control modules by aggregating their latent variables between diffusion model blocks. The second finding enables a multi-inference module reweighting strategy that balances module contributions to generation, requiring no additional training or fine-tuning overhead. Extensive results demonstrate that ModuleTeam not only outperforms existing methods but also provides flexibility in the types of conditions and scalability in their number.

Customized text-to-video generation aims to generate text-guided videos with user-given subjects, which has gained increasing attention. However, existing works are primarily limited to single-subject oriented text-to-video generation, leaving the more challenging problem of customized multi-subject generation unexplored. In this paper, we fill this gap and propose a novel VideoDreamer framework, which can generate temporally consistent text-guided videos that faithfully preserve the visual features of the given multiple subjects. Specifically, VideoDreamer adopts the pretrained Stable Diffusion with temporal modules as its base video generator, taking the power of the text-to-image model to generate diversified content. The video generator is further customized for multi-subjects, which leverages the proposed Disen-Mix Finetuning and Human-in-the-Loop Re-finetuning strategy, to tackle the attribute binding problem of multi-subject generation. Additionally, we present a disentangled motion customization strategy to finetune the temporal modules so that we can generate videos with both customized subjects and motions. To evaluate the performance of customized multi-subject text-to-video generation, we introduce the MultiStudioBench benchmark. Extensive experiments demonstrate the remarkable ability of VideoDreamer to generate videos with new content such as new events and backgrounds, tailored to the customized multiple subjects.

Sequential recommendation aims to recommend the next items that a target user may have interest in based on the user’s sequence of past behaviors, which has become a hot research topic in both academia and industry. In the literature, sequential recommendation adopts a Sequence-to-Item or Sequence-to-Sequence training strategy, which supervises a sequential model with a user’s next one or more behaviors as the labels and the sequence of the past behaviors as the input. However, existing powerful sequential recommendation approaches employ more and more complex deep structures such as Transformer in order to accurately capture the sequential patterns, which heavily rely on hand-crafted designs on key attention mechanism to achieve state-of-the-art performance, thus failing to automatically obtain the optimal design of attention representation architectures in various scenarios with different data. Other works on classic automated deep recommender systems only focus on traditional settings, ignoring the problem of sequential scenarios. In this article, we study the problem of automated sequential recommendation, which faces two main challenges: (1) How can we design a proper search space tailored for attention automation in sequential recommendation, and (2) How can we accurately search effective attention representation architectures considering multiple user interests reflected in the sequential behavior. To tackle these challenges, we propose an automated disentangled sequential recommendation (AutoDisenSeq) model. In particular, we employ neural architecture search (NAS) and design a search space tailored for automated attention representation in attentive intention-disentangled sequential recommendation with an expressive and efficient space complexity of \(O(n^{2})\) given \(n\) as the number of layers. We further propose a context-aware parameter sharing mechanism taking characteristics of each sub-architecture into account to enable accurate architecture performance estimations and great flexibility for disentanglement of latent intention representation. Moreover, we propose AutoDisenSeq-large language model (LLM), which utilizes the textual understanding power of LLM as a guidance to refine the candidate list for recommendation from AutoDisenSeq. We conduct extensive experiments to show that our proposed AutoDisenSeq model and AutoDisenSeq-LLM model outperform existing baseline methods on four real-world datasets in both overall recommendation and cold-start recommendation scenarios.

Large language models (LLMs) have shown remarkable text understanding capabilities, which have been ex-tended as Video LLMs to handle video data for compre-hending visual details. However, existing Video LLMs can only provide a coarse description of the entire video, failing to capture the precise start and end time bound-ary of specific events. In this paper, we solve this issue via proposing VTimeLLM, a novel Video LLM designed for fine-grained video moment understanding and reasoning with respect to time boundary. Specifically, our VTimeLLM adopts a boundary-aware three-stage training strategy, which respectively utilizes image-text pairs for feature alignment, multiple-event videos to increase temporal-boundary awareness, and high-quality video-instruction tuning to further improve temporal understanding ability as well as align with human intents. Extensive experiments demonstrate that in fine-grained time-related comprehension tasks for videos such as Temporal Video Grounding and Dense Video Captioning, VTimeLLM significantly outperforms existing Video LLMs. Besides, benefits from the fine-grained temporal understanding of the videos further enable VTimeLLM to beat existing Video LLMs in video di-alogue benchmark, showing its superior cross-modal understanding and reasoning abilities.11Our project page is at https://github.com/huangb23/VTimeLLM

Subject-driven text-to-image generation aims to generate customized images of the given subject based on the text descriptions, which has drawn increasing attention. Existing methods mainly resort to finetuning a pretrained generative model, where the identity-relevant information (e.g., the boy) and the identity-irrelevant information (e.g., the background or the pose of the boy) are entangled in the latent embedding space. However, the highly entangled latent embedding may lead to the failure of subject-driven text-to-image generation as follows: (i) the identity-irrelevant information hidden in the entangled embedding may dominate the generation process, resulting in the generated images heavily dependent on the irrelevant information while ignoring the given text descriptions; (ii) the identity-relevant information carried in the entangled embedding can not be appropriately preserved, resulting in identity change of the subject in the generated images. To tackle the problems, we propose DisenBooth, an identity-preserving disentangled tuning framework for subject-driven text-to-image generation. Specifically, DisenBooth finetunes the pretrained diffusion model in the denoising process. Different from previous works that utilize an entangled embedding to denoise each image, DisenBooth instead utilizes disentangled embeddings to respectively preserve the subject identity and capture the identity-irrelevant information. We further design the novel weak denoising and contrastive embedding auxiliary tuning objectives to achieve the disentanglement. Extensive experiments show that our proposed DisenBooth framework outperforms baseline models for subject-driven text-to-image generation with the identity-preserved embedding. Additionally, by combining the identity-preserved embedding and identity-irrelevant embedding, DisenBooth demonstrates more generation flexibility and controllability.

The existing graph neural architecture search (GNAS) methods assume that the graph tasks are static during the search process, ignoring the ubiquitous scenarios where sequential graph tasks come in a continual fashion. Moreover, existing GNAS works resort to entangled graph factors during the architecture search process, resulting in the catastrophic forgetting problems. In this paper, we study the problem of continual graph neural architecture search that is expected to continually search the architecture to learn new graph tasks without forgetting the past, which remains largely unexplored in the literature. However, this problem poses the challenge of architecture conflicts, i.e., the optimal architecture for the new graph task may have performance deterioration and thus sub-optimal for past tasks. To address the challenge, we propose a novel Disentangled Continual Graph Neural Architecture Search with Invariant Modularization (GASIM) method, which is able to continually search the optimal architectures without forgetting past knowledge. Specifically, we first design a modular graph architecture super-network incorporating multiple modules to enable searching architecture with factor expertise. Second, we propose a factor-based task-module router that discovers the latent graph factors and routes the incoming task to the best suitable architecture module to alleviate the forgetting problem induced by architecture conflicts. Finally, we propose an invariant architecture search mechanism to capture the shared knowledge among tasks. Extensive experiments on real-world datasets demonstrate that the proposed method achieves state-of-the-art performance against baselines in continual graph neural architecture search.

Curriculum learning is a training paradigm where machine learning models are trained in a meaningful order, inspired by the way humans learn curricula. Due to its capability to improve model generalization and convergence, curriculum learning has gained considerable attention and has been widely applied to various research domains. Nevertheless, as new curriculum learning methods continue to emerge, it remains an open issue to benchmark them fairly. Therefore, we develop CurBench, the first benchmark that supports systematic evaluations for curriculum learning. Specifically, it consists of 15 datasets spanning 3 research domains: computer vision, natural language processing, and graph machine learning, along with 3 settings: standard, noise, and imbalance. To facilitate a comprehensive comparison, we establish the evaluation from 2 dimensions: performance and complexity. CurBench also provides a unified toolkit that plugs automatic curricula into general machine learning processes, enabling the implementation of 15 core curriculum learning methods. On the basis of this benchmark, we conduct comparative experiments and make empirical analyses of existing methods. CurBench is open-source and publicly available at https://github.com/THUMNLab/CurBench.

Visual concept recognition aims to capture the basic attributes of an image and reason about the relationships among them to determine whether the image satisfies a certain concept, and has been widely used in various tasks such as human action recognition and image risk warning. Most existing works adopt deep neural networks for visual concept recognition, which are black-box and incomprehensible to humans, thus making them unacceptable for sensitive domains such as prohibited event detection and risk early warning etc. To address this issue, we propose to combine large language model (LLM) with explainable symbolic reasoning via curriculum reweighting to increase the interpretability and accuracy of visual concept recognition in this paper. However, realizing this goal is challenging given that i) the performance of symbolic representations are limited by the lack of annotated reasoning symbols and rules for most tasks, and ii) the LLMs may suffer from knowlege hallucination and dynamic open environment. To address these issues, in this paper, we propose CurLLM-Reasoner, a curriculum reasoning method based on symbolic reasoning and large language model for visual concept recognition. Specifically, we propose a novel rule enhancement module with a tool library, which fully leverage the reasoning capability of large language models and can generate human-understandable rules without any annotation. We further propose a curriculum data resampling methodology to help the large language model accurately extract from easy to complex rules at different reasoning stages. Extensive experiments on various datasets demonstrate that CurLLM-Reasoner can achieve the state-of-the-art visual concept recognition results with explainable rules while free of human annotations.

This tutorial focuses on curriculum learning (CL), an important topic in machine learning, which gains an increasing amount of attention in the research community. CL is a learning paradigm that enables machines to learn from easy data to hard data, imitating the meaningful procedure of human learning with curricula. As an easy-to-use plug-in, CL has demonstrated its power in improving the generalization capacity and convergence rate of various models in a wide range of scenarios such as computer vision, natural language processing, reinforcement learning, etc. In particular, CL can also play an important role in multimedia applications. Therefore, it is essential to introduce CL to more scholars and researchers in the machine learning and multimedia community. However, there have been no tutorials on CL for multimedia so far, motivating the organization of this tutorial at ACM Multimedia 2024. To give a comprehensive tutorial on CL for multimedia, we plan to organize it from the following aspects: (1) theories, (2) approaches, (3) applications, (4) tools, and (5) future directions. First, we introduce the motivations, theories, and insights behind CL. Second, we advocate novel, high-quality approaches, as well as innovative solutions to the challenging problems in CL. Then we present the applications of CL in various scenarios, especially multimedia, followed by some relevant tools. In the end, we discuss open questions and future directions in the era of large language models. We believe this topic is at the core of the scope of ACM Multimedia and is attractive to the audience interested in machine learning and multimedia from both academia and industry.

Generating customized content in videos has received increasing attention recently. However, existing works primarily focus on customized text-to-video generation for single subject, suffering from subject-missing and attribute-binding problems when the video is expected to contain multiple subjects. Furthermore, existing models struggle to assign the desired actions to the corresponding subjects (action-binding problem), failing to achieve satisfactory multi-subject generation performance. To tackle the problems, in this paper, we propose DisenStudio, a novel framework that can generate text-guided videos for customized multiple subjects, given few images for each subject. Specifically, DisenStudio enhances a pretrained diffusion-based text-to-video model with our proposed spatial-disentangled cross-attention mechanism to associate each subject with the desired action. Then the model is customized for the multiple subjects with the proposed motion-preserved disentangled finetuning, which involves three tuning strategies: multi-subject co-occurrence tuning, masked single-subject tuning, and multi-subject motion-preserved tuning. The first two strategies guarantee the subject occurrence and preserve their visual attributes, and the third strategy helps the model maintain the temporal motion-generation ability when finetuning on static images. We conduct extensive experiments to demonstrate our proposed DisenStudio significantly outperforms existing methods in various metrics. Additionally, we show that DisenStudio can be used as a powerful tool for various controllable generation applications.

Sampling strategies have been widely adopted in Vision-Language Pre-training (VLP) and have achieved great success recently. However, the sampling strategies adopted by current VLP works are limited in two ways: i) they only focus on negative sampling, ignoring the importance of more informative positive samples; ii) their sampling strategies are conducted in the local in-batch level, which may lead to sub-optimal results. To tackle these problems, in this paper, we propose a curriculum-based Global Positive-Negative Sampling (GPN-S) framework for vision-language pre-training, which conducts both positive and negative sampling in the global level, grounded on the notion of neighborhood relationships. Additionally, we incorporate curriculum learning into our sampling strategy, progressively increasing the complexity of samples as the training progresses. Specifically, our proposed GPN-S framework is capable of utilizing positive sampling to bring semantically equivalent samples closer, as well as employing negative sampling to push challenging negative samples farther away. We jointly consider them for vision-language pre-training on the global-level perspective rather than a local-level mini-batch, which provides more informative and diverse samples. We evaluate the effectiveness of the proposed GPN-S framework by conducting experiments on several common downstream tasks, and the results demonstrate significant performance improvement over the existing models.

Existing Video Corpus Moment Retrieval (VCMR) is limited to coarse-grained understanding that hinders precise video moment localization when given fine-grained queries. In this paper, we propose a more challenging fine-grained VCMR benchmark requiring methods to localize the best-matched moment from the corpus with other partially matched candidates. To improve the dataset construction efficiency and guarantee high-quality data annotations, we propose VERIFIED, an automatic \underline{V}id\underline{E}o-text annotation pipeline to generate captions with \underline{R}el\underline{I}able \underline{FI}n\underline{E}-grained statics and \underline{D}ynamics. Specifically, we resort to large language models (LLM) and large multimodal models (LMM) with our proposed Statics and Dynamics Enhanced Captioning modules to generate diverse fine-grained captions for each video. To filter out the inaccurate annotations caused by the LLM hallucination, we propose a Fine-Granularity Aware Noise Evaluator where we fine-tune a video foundation model with disturbed hard-negatives augmented contrastive and matching losses. With VERIFIED, we construct a more challenging fine-grained VCMR benchmark containing Charades-FIG, DiDeMo-FIG, and ActivityNet-FIG which demonstrate a high level of annotation quality. We evaluate several state-of-the-art VCMR models on the proposed dataset, revealing that there is still significant scope for fine-grained video understanding in VCMR.

Disentangled Representation Learning (DRL) aims to learn a model capable of identifying and disentangling the underlying factors hidden in the observable data in representation form. The process of separating underlying factors of variation into variables with semantic meaning benefits in learning explainable representations of data, which imitates the meaningful understanding process of humans when observing an object or relation. As a general learning strategy, DRL has demonstrated its power in improving the model explainability, controlability, robustness, as well as generalization capacity in a wide range of scenarios such as computer vision, natural language processing, and data mining. In this article, we comprehensively investigate DRL from various aspects including motivations, definitions, methodologies, evaluations, applications, and model designs. We first present two well-recognized definitions, i.e., Intuitive Definition and Group Theory Definition for disentangled representation learning. We further categorize the methodologies for DRL into four groups from the following perspectives, the model type, representation structure, supervision signal, and independence assumption. We also analyze principles to design different DRL models that may benefit different tasks in practical applications. Finally, we point out challenges in DRL as well as potential research directions deserving future investigations. We believe this work may provide insights for promoting the DRL research in the community.

Subject-driven text-to-image generation aims to generate customized images of the given subject based on the text descriptions, which has drawn increasing attention recently. Existing methods mainly resort to finetuning a pretrained generative model, where the identity-relevant information (e.g., the boy) and the identity-irrelevant sample-specific information (e.g., the background or the pose of the boy) are entangled in the latent embedding space. However, the highly entangled latent embedding may lead to low subject identity fidelity and text prompt fidelity. To tackle the problems, we propose DisenDreamer, a sample-aware disentangled tuning framework for subject-driven text-to-image generation in this paper. Specifically, DisenDreamer finetunes the pretrained diffusion model in the denoising process. Different from previous works that utilize an entangled embedding to denoise, DisenDreamer instead utilizes a common text embedding to capture the identity-relevant information and a sample-specific visual embedding to capture the identity-irrelevant information. To disentangle the two embeddings, we further design the novel weak common denoising, weak sample-aware denoising, and the contrastive embedding auxiliary tuning objectives. Extensive experiments show that our proposed DisenDreamer framework outperforms baseline models for subject-driven text-to-image generation. Additionally, by combining the identity-relevant and the identity-irrelevant embedding, DisenDreamer demonstrates more generation flexibility and controllability.

Video question answering (VideoQA) requires the ability of comprehensively understanding visual contents in videos. Existing VideoQA models mainly focus on scenarios involving a single event with simple object interactions and leave event-centric scenarios involving multiple events with dynamically complex object interactions largely unexplored. These conventional VideoQA models are usually based on features extracted from the global visual signals, making it difficult to capture the object-level and event-level semantics. Although there exists a recent work utilizing a static spatio-temporal graph to explicitly model object interactions in videos, it ignores the dynamic impact of questions for graph construction and fails to exploit the implicit event-level semantic clues in questions. To overcome these limitations, we propose a Self-supervised Dynamic Graph Reasoning (SDGraphR) model for video question answering (VideoQA). Our SDGraphR model learns a question-guided spatio-temporal graph that dynamically encodes intra-frame spatial correlations and inter-frame correspondences between objects in the videos. Furthermore, the proposed SDGraphR model discovers event-level cues from questions to conduct self-supervised learning with an auxiliary event recognition task, which in turn helps to improve its VideoQA performances without using any extra annotations. We carry out extensive experiments to validate the substantial improvements of our proposed SDGraphR model over existing baselines.

Video Grounding (VG) aims to locate the desired segment from a video given a sentence query. Recent studies have found that current VG models are prone to over-rely the groundtruth moment annotation distribution biases in the training set. To discourage the standard VG model's behavior of exploiting such temporal annotation biases and improve the model generalization ability, we propose multiple negative augmentations in a hierarchical way, including cross-video augmentations from clip-/video-level, and self-shuffled augmentations with masks. These augmentations can effectively diversify the data distribution so that the model can make more reasonable predictions instead of merely fitting the temporal biases. However, directly adopting such data augmentation strategy may inevitably carry some noise shown in our cases, since not all of the handcrafted augmentations are semantically irrelevant to the groundtruth video. To further denoise and improve the grounding accuracy, we design a multi-stage curriculum strategy to adaptively train the standard VG model from easy to hard negative augmentations. Experiments on newly collected Charades-CD and ActivityNet-CD datasets demonstrate our proposed strategy can improve the performance of the base model on both i.i.d and o.o.d scenarios.

There exist complex patterns behind the decision-making processes of different individuals across different environments. For instance, in a social recommender system, various user behaviors are driven by highly entangled latent factors from two environments, i.e., consuming environment where users consume items and social environment where users connect with each other. Uncovering the disentanglement of these latent factors for users can benefit in enhanced explainability and controllability for recommendation. However, in literature there has been no work on social recommendation capable of disentangling user representations across consuming and social environments. To solve this problem, we study co-disentangled representation learning across different environments via proposing the curriculum co-disentangled representation learning (CurCoDis) model to disentangle the hidden factors for users across both consuming and social environments. To co-disentangle joint representations for user-item consumption and user-user social graph simultaneously, we partition the social graph into equal-size sub-graphs with minimum number of edges being cut, and design a curriculum weighing strategy for subgraph training through measuring the complexity of subgraphs via Descartes' rule of signs. We further develop the prototype-routing optimization mechanism, which achieves co-disentanglement of user representations across consuming and social environments. Extensive experiments for social recommendation demonstrate that our proposed CurCoDis model can significantly outperform state-of-the-art methods on several real-world datasets.

Sequential recommendation aims at mining time-aware user interests through modeling sequential behaviors. Transformer, as an effective architecture designed to process sequential input data, has shown its superiority in capturing sequential relations for recommendation. Nevertheless, existing Transformer architectures lack explicit regularization for layer-wise disentanglement, which fails to take advantage of disentangled representation in recommendation and leads to suboptimal performance. In this paper, we study the problem of layer-wise disentanglement for Transformer architectures and propose the Adaptive Disentangled Transformer (ADT) framework, which is able to adaptively determine the optimal degree of disentanglement of attention heads within different layers. Concretely, we propose to encourage disentanglement by requiring the independence constraint via mutual information estimation over attention heads and employing auxiliary objectives to prevent the information from collapsing into useless noise. We further propose a progressive scheduler to adaptively adjust the weights controlling the degree of disentanglement via an evolutionary process. Extensive experiments on various real-world datasets demonstrate the effectiveness of our proposed ADT framework.

Disentangled Representation Learning (DRL) aims to learn a model capable of identifying and disentangling the underlying factors hidden in the observable data in representation form. The process of separating underlying factors of variation into variables with semantic meaning benefits in learning explainable representations of data, which imitates the meaningful understanding process of humans when observing an object or relation. As a general learning strategy, DRL has demonstrated its power in improving the model explainability, controllability, robustness, as well as generalization capacity in a wide range of scenarios such as computer vision, natural language processing, data mining etc. In this tutorial, we comprehensively present DRL from various aspects including motivations, definitions, methodologies, evaluations, applications and model designs for multimedia. We discuss works on DRL based on two well-recognized definitions, i.e., Intuitive Definition and Group Theory Definition. We further categorize the methodologies for DRL into four groups, i.e., Traditional Statistical Approaches, Variational Auto-encoder Based Approaches, Generative Adversarial Networks Based Approaches, Hierarchical Approaches and Other Approaches. We also analyze principles to design different DRL models that may benefit different tasks in practical multimedia applications. Finally, we point out challenges in DRL as well as potential research directions deserving future investigations. We believe this tutorial may provide insights for promoting the DRL research in the multimedia community.

Video Grounding (VG), has drawn widespread attention over the past few years, and numerous studies have been devoted to improving performance on various VG benchmarks. Nevertheless, the label annotation procedures in VG produce imbalanced query-moment-label distributions in the datasets, which severely deteriorate the learning model's capability of truly understanding the video contents. Existing works on debiased VG either focus on adjusting the learning model or conducting video-level augmentation, failing to handle the temporal bias issue caused by imbalanced query-moment-label distributions. In this paper, we propose a Disentangled Feature Mixup (DFM) framework for debiased VG, which is capable of performing unbiased grounding to tackle the temporal bias issue. Specifically, a feature-mixup augmentation strategy is designed to generate new (text, location) pairs with diverse temporal distributions via jointly augmenting the representation of text queries and the location labels. This strategy encourages making prediction based on more diverse data samples with balanced query-moment-label distributions. Furthermore, we also design a content-location disentanglement module to disentangle the representations of the temporal information and content information in videos, which is able to remove the spurious effect of temporal biases on video representation. Given that our proposed DFM framework conducts feature-level augmentation and disentanglement, it is model-agnostic and can be applied to most baselines simply yet effectively. Extensive experiments show that our proposed DFM framework is able to significantly outperform baseline models in various metrics under both independent identical distribution (i.i.d.) and out-of-distribution (o.o.d.) scenes, especially in scenarios with annotation distribution changes.

Temporal Sentence Grounding aims to retrieve a video moment given a natural language query. Most existing literature merely focuses on visual information in videos without considering the naturally accompanied audio which may contain rich semantics. The few works considering audio simply regard it as an additional modality, overlooking that: i) it's non-trivial to explore consistency and complementarity between audio and visual; ii) such exploration requires handling different levels of information densities and noises in the two modalities. To tackle these challenges, we propose Adaptive Dual-branch Promoted Network (ADPN) to exploit such consistency and complementarity: i) we introduce a dual-branch pipeline capable of jointly training visual-only and audio-visual branches to simultaneously eliminate inter-modal interference; ii) we design Text-Guided Clues Miner (TGCM) to discover crucial locating clues via considering both consistency and complementarity during audio-visual interaction guided by text semantics; iii) we propose a novel curriculum-based denoising optimization strategy, where we adaptively evaluate sample difficulty as a measure of noise intensity in a self-aware fashion. Extensive experiments show the state-of-the-art performance of our method.

Knowledge graphs serve as a powerful tool to boost model performances for various applications covering computer vision, natural language processing, multimedia data mining, etc. The process of knowledge acquisition for human is multimodal in essence, covering text, image, video and audio modalities. However, existing multimodal knowledge graphs fail to cover all these four elements simultaneously, severely limiting their expressive powers in performance improvement for downstream tasks. In this paper, we propose TIVA-KG, a multimodal Knowledge Graph covering Text, Image, Video and Audio, which can benefit various downstream tasks. Our proposed TIVA-KG has two significant advantages over existing knowledge graphs in i) coverage of up to four modalities including text, image, video, audio, and ii) capability of triplet grounding which grounds multimodal relations to triples instead of entities. We further design a Quadruple Embedding Baseline (QEB) model to validate the necessity and efficacy of considering four modalities in KG. We conduct extensive experiments to test the proposed TIVA-KG with various knowledge graph representation approaches over link prediction task, demonstrating the benefits and necessity of introducing multiple modalities and triplet grounding. TIVA-KG is expected to promote further research on mining multimodal knowledge graph as well as the relevant downstream tasks in the community. TIVA-KG is now available at our website: http://mn.cs.tsinghua.edu.cn/tivakg.

Sequential recommender systems often suffer from performance drops due to the data-sparsity issue in real-world scenarios. To address this issue, we bravely take advantage of the strength in diffusion model to conduct data augmentation for sequential recommendation in this paper. However, there remain two critical challenges for this scarcely-explored topic: (i) previous diffusion models are mostly designed for image generation aiming to capture pixel patterns, which can hardly be applied in data augmentation for sequential recommendation aiming to capture the user-item relations; (ii) given a specific diffusion model capable of user-item interaction augmentation, it is non-trivial to guarantee that the diffusion-generated data can always bring benefits towards the sequential recommendation model. To tackle these challenges, we propose Diff4Rec, a curriculum-scheduled diffusion augmentation framework for sequential recommendation. Specifically, a diffusion model is pre-trained on recommendation data via corrupting and reconstructing the user-item interactions in the latent space, and the generated predictions are leveraged to produce diversified augmentations for the sparse user-item interactions. Subsequently, a curriculum scheduling strategy is designed to progressively feed the diffusion-generated samples into the sequential recommenders, with respect to two levels, i.e., interaction augmentation and objective augmentation, to jointly optimize the data and model. Extensive experiments demonstrate that our proposed Diff4Rec framework is able to effectively achieve superior performance over several strong baselines, capable of making high-quality and robust sequential recommendations. We believe the proposed Diff4Rec has the promising potential to bring paradigm shift in multimedia recommendation.

Multimodal learning has been widely studied and applied due to its improvement over previous unimodal tasks and its effectiveness on emerging multimodal challenges. However, it has been reported that modal encoders are under-optimized in multimodal learning in contrast to unimodal learning, especially when some modalities are dominant over others. Existing solutions to this problem suffer from two limitations: i) they merely focus on inter-modal balance, failing to consider the influence of intra-modal data on each modality; ii) their implementations heavily rely on unimodal performances or losses, thus being suboptimal for the tasks requiring modal interactions (e.g., visual question answering). To tackle these limitations, we propose I2MCL, a generic Intra- and Inter-Modal Curriculum Learning framework which simultaneously considers both data difficulty and modality balance for multimodal learning. In the intra-modal curriculum, we adopt a pretrained teacher model to obtain knowledge distillation loss as the difficulty measurer, which determines the data weights within the corresponding modality. In the inter-modal curriculum, we utilize a Pareto optimization strategy to measure and compare the gradients from distillation loss and task loss across modalities, capable of determining whether a modality should learn from the task or its teacher. Empirical experiments on various tasks including multimodal classification, visual question answering and visual entailment demonstrate that our proposed I2MCL is able to tackle the under-optimized modality problem and bring consistent improvement to multimodal learning.

Transformer-based model has gained great success in the multimedia sequential recommendation task due to its strong ability to handle sequential data. However, existing Transformer-based models regard the items in the sequential data as a user-specific fully-connected graph (local graph) and only explicitly consider the temporal information in the local graph to capture the users’ intentions, ignoring the fact that the user-item bipartite graph (global graph) may carry important relation patterns to the sequential items. Additionally, it is still unclear whether (and how) the information hidden in the global graphs can help the Transformer-based models better understand the users’ sequential behavior according to the current literature. To investigate this important problem, we propose to utilize the global graph information to help the Transformer-based sequential recommendation, where the information from different modalities, i.e., user-item interactions in the global graph and the temporal patterns in the historical sequences, are taken into account jointly. In concrete, we propose two Global-Local (GL) GraphFormer models for utilizing both the global graph and local temporal information. One GL-GraphFormer is able to gift the Transformer-based model with both first- and second-order graph information through two specifically designed encodings. The other GL-GraphFormer transfers higher-order graph information into the local Transformer with pretrained Graph Neural Networks (GNNs). Extensive experiments on several real-world datasets demonstrate that i) our proposed GL-GraphFormers can bring substantial improvement over baseline methods, and ii) the benefits of different orders of global graph information vary with the dataset sparsity.

Current auxiliary learning methods mainly adopt the methodology of reweighing losses for the manually collected auxiliary data and tasks. However, these methods heavily rely on domain knowledge during data collection, which may be hardly available in reality. Therefore, current methods will become less effective and even do harm to the primary task when unhelpful auxiliary data and tasks are employed. To tackle the problem, we propose a joint data-task generation framework for auxiliary learning (DTG-AuxL), which can bring benefits to the primary task by generating the new auxiliary data and task in a joint manner. The proposed DTG-AuxL framework contains a joint generator and a bi-level optimization strategy. Specifically, the joint generator contains a feature generator and a label generator, which are designed to be applicable and expressive for various auxiliary learning scenarios. The bi-level optimization strategy optimizes the joint generator and the task learning model, where the joint generator is effectively optimized in the upper level via the implicit gradient from the primary loss and the explicit gradient of our proposed instance regularization, while the task learning model is optimized in the lower level by the generated data and task. Extensive experiments show that our proposed DTG-AuxL framework consistently outperforms existing methods in various auxiliary learning scenarios, particularly when the manually collected auxiliary data and tasks are unhelpful.

Graph neural architecture search (GraphNAS) has shown great potential for automatically designing graph neural architectures for graph related tasks. However, multi-task GraphNAS capable of handling multiple tasks simultaneously has been largely unexplored in literature, posing great challenges to capture the complex relations and influences among different tasks. To tackle this problem, we propose a novel multi-task graph neural architecture search with task-aware collaboration and curriculum (MTGC3), which is able to simultaneously discover optimal architectures for different tasks and learn the collaborative relationships among different tasks in a joint manner. Specifically, we design the layer-wise disentangled supernet capable of managing multiple architectures in a unified framework, which combines with our proposed soft task-collaborative module to learn the transferability relationships between tasks. We further develop the task-wise curriculum training strategy to improve the architecture search procedure via reweighing the influence of different tasks based on task difficulties. Extensive experiments show that our proposed MTGC3 model achieves state-of-the-art performance against several baselines in multi-task scenarios, demonstrating its ability to discover effective architectures and capture the collaborative relationships for multiple tasks.

Cross-domain recommendation (CDR) aims to leverage the source domain information to provide better recommendation for the target domain, which is widely adopted in recommender systems to alleviate the data sparsity and cold-start problems. However, existing CDR methods mostly focus on designing effective model architectures to transfer the source domain knowledge, ignoring the behavior-level effect during the loss optimization process, where behaviors regarding different aspects in the source domain may have different importance for the CDR model optimization. The ignorance of the behavior-level effect will cause the carefully designed model architectures ending up with sub-optimal parameters, which limits the recommendation performance. To tackle the problem, we propose a generic behavioral importance-aware optimization framework for cross-domain recommendation (BIAO). Specifically, we propose a behavioral perceptron which predicts the importance of each source behavior according to the corresponding item’s global impact and local user-specific impact. The joint optimization process of the CDR model and the behavioral perceptron is formulated as a bi-level optimization problem. In the lower optimization, only the CDR model is updated with weighted source behavior loss and the target domain loss, while in the upper optimization, the behavioral perceptron is updated with implicit gradient from a developing dataset obtained through the proposed reorder-and-reuse strategy. Extensive experiments show that our proposed optimization framework consistently improves the performance of different cross-domain recommendation models in 7 cross-domain scenarios, demonstrating that our method can serve as a generic and powerful tool for cross-domain recommendation1.

There exist complex interactions among a large number of latent factors behind the decision making processes of different individuals, which drive the various user behavior patterns in recommender systems. These factors hidden in those diverse behaviors demonstrate highly entangled patterns, covering from high-level user intentions to low-level individual preferences. Uncovering the disentanglement of these latent factors can benefit in enhanced robustness, interpretability, and controllability during representation learning for recommendation. However, the large degree of entanglement within latent factors poses great challenges for learning representations that disentangle them, and remains largely unexplored in literature. In this paper, we present the SEMantic MACRo-mIcro Disentangled Variational Auto-Encoder (SEM-MacridVAE) model for learning disentangled representations from user behaviors, taking item semantic information into account. Our SEM-MacridVAE model achieves macro disentanglement by inferring the high-level concepts associated with user intentions (e.g., to buy a pair of shoes or a laptop) through a prototype routing mechanism, as well as capturing the individual preferences with respect to different concepts separately. The micro disentanglement is guaranteed through a micro-disentanglement regularizer stemming from an information-theoretic interpretation of VAEs, which forces each dimension of the representations to independently reflect an isolated low-level factor (e.g., the size or the color of a shirt). The semantic information including visual and categorical signals extracted from candidate items is utilized to further boost the recommendation performance of the proposed SEM-MacridVAE model. Empirical experiments demonstrate that our proposed approach is able to achieve significant improvement over the state-of-the-art baselines. We also show that the learned representations are interpretable and controllable, capable of potentially leading to a new paradigm for recommendation where users have fine-grained control over some target aspects of the recommendation candidates.

Existing auxiliary learning approaches only consider the relationships between the target task and the auxiliary tasks, ignoring the fact that data samples within an auxiliary task could contribute differently to the target task, which results in inefficient auxiliary information usage and non-robustness to data noise. In this paper, we propose to learn a joint task and data schedule for auxiliary learning, which captures the importance of different data samples in each auxiliary task to the target task. However, learning such a joint schedule is challenging due to the large number of additional parameters required for the schedule. To tackle the challenge, we propose a joint task and data scheduling (JTDS) model for auxiliary learning. The JTDS model captures the joint task-data importance through a task-data scheduler, which creates a mapping from task, feature and label information to the schedule in a parameter-efficient way. Particularly, we formulate the scheduler and the task learning process as a bi-level optimization problem. In the lower optimization, the task learning model is updated with the scheduled gradient, while in the upper optimization, the task-data scheduler is updated with the implicit gradient. Experimental results show that our JTDS model significantly outperforms the state-of-the-art methods under supervised, semi-supervised and corrupted label settings.

Graph Neural Architecture Search (GNAS) has become a powerful method in automatically discovering suitable Graph Neural Network (GNN) architectures for different tasks. However, existing approaches fail to handle large-scale graphs because current performance estimation strategies in GNAS are computationally expensive for large-scale graphs and suffer from consistency collapse issues. To tackle these problems, we propose the Graph ArchitectUre Search at Scale (GAUSS) method that can handle large-scale graphs by designing an efficient light-weight supernet and the joint architecture-graph sampling. In particular, a graph sampling-based single-path one-shot supernet is proposed to reduce the computation burden. To address the consistency collapse issues, we further explicitly consider the joint architecture-graph sampling through a novel architecture peer learning mechanism on the sampled sub-graphs and an architecture importance sampling algorithm. Our proposed framework is able to smooth the highly non-convex optimization objective and stabilize the architecture sampling process. We provide theoretical analyses on GAUSS and empirically evaluate it on five datasets whose vertex sizes range from 10^4 to 10^8. The experimental results demonstrate substantial improvements of GAUSS over other GNAS baselines on all datasets. To the best of our knowledge, the proposed GAUSS method is the first graph neural architecture search framework that can handle graphs with billions of edges within 1 GPU day.

Video Question Answering (VideoQA) aims to understand given videos and questions comprehensively by generating correct answers. However, existing methods usually rely on end-to-end black-box deep neural networks to infer the answers, which significantly differs from human logic reasoning, thus lacking the ability to explain. Besides, the performances of existing methods tend to drop when answering compositional questions involving realistic scenarios. To tackle these challenges, we propose a Dynamic Spatio-Temporal Modular Network (DSTN) model, which utilizes a spatio-temporal modular network to simulate the compositional reasoning procedure of human beings. Concretely, we divide the task of answering a given question into a set of sub-tasks focusing on certain key concepts in questions and videos such as objects, actions, temporal orders, etc. Each sub-task can be solved with a separately designed module, e.g., spatial attention module, temporal attention module, logic module, and answer module. Then we dynamically assemble different modules assigned with different sub-tasks to generate a tree-structured spatio-temporal modular neural network for human-like reasoning before producing the final answer for the question. We carry out extensive experiments on the AGQA dataset to demonstrate our proposed DSTN model can significantly outperform several baseline methods in various settings. Moreover, we evaluate intermediate results and visualize each reasoning step to verify the rationality of different modules and the explainability of the proposed DSTN model.

Audio-visual question answering aims to answer questions regarding both audio and visual modalities in a given video, and has drawn increasing research interest in recent years. However, there have been no appropriate datasets for this challenging task on videos in real-life scenarios so far. They are either designed with questions containing only visual clues without taking any audio information into account, or considering audio with restrictions to specific scenarios, such as panoramic videos and videos about music performances. In this paper, to overcome the limitations of existing datasets, we introduce AVQA, a new audio-visual question answering dataset on videos in real-life scenarios. We collect 57,015 videos from daily audio-visual activities and 57,335 specially-designed question-answer pairs relying on clues from both modalities, where information contained in a single modality is insufficient or ambiguous. Furthermore, we propose a Hierarchical Audio-Visual Fusing module to model multiple semantic correlations among audio, visual, and text modalities and conduct ablation studies to analyze the role of different modalities on our datasets. Experimental results show that our proposed method significantly improves the audio-visual question answering performance over various question types. Therefore, AVQA can provide an adequate testbed for the generation of models with a deeper understanding of multimodal information on audio-visual question answering in real-life scenarios. (The dataset is available at https://mn.cs.tsinghua.edu.cn/avqa)

Neural Architecture Search (NAS) is an effective way to automatically design neural architectures for various multimedia applications. Weight-sharing, as one of the most popular NAS strategies, has been widely adopted due to its search efficiency. Existing weight-sharing NAS methods overlook the influence of data distribution and treat each data sample equally. Contrastively, in this paper, we empirically discover that different data samples have different influences on architectures, e.g., some data samples are easy to fit by certain architectures but hard by others. Hence, there exist architectures with better performances on early data samples being more likely to be discovered in the whole NAS searching process, which leads to a suboptimal searching result. To tackle this problem, we propose Curriculum-NAS, a curriculum training framework on weight-sharing NAS, which dynamically changes the training data weights during the searching process. In particular, Curriculum-NAS utilizes the multiple subnets included in weight-sharing NAS to jointly assess data uncertainty, which serves as the difficulty criterion in a curriculum manner, so that the potentially optimal architectures can obtain higher probability of being fully trained and discovered. Extensive experiments on several image and text datasets demonstrate that our Curriculum-NAS can bring consistent improvement over existing weight-sharing NAS. The code is available online at https://github.com/zhouyw16/curriculum-nas.

Curriculum learning (CL) is a machine learning paradigm gradually learning from easy to hard, which is inspired by human curricula. As an easy-to-use and general training strategy, CL has been widely applied to various multimedia tasks covering images, texts, audios, videos, etc. The effectiveness of CL has recently facilitated an increasing number of new CL algorithms. However, there has been no open-source library for curriculum learning, making it hard to reproduce, evaluate and compare the numerous CL algorithms on fair benchmarks and settings. To ease and promote future research on CL, we develop CurML, the first Curriculum Machine L earning library to integrate existing CL algorithms into a unified framework. It is convenient to use and flexible to customize by calling the provided five APIs, which are designed for easily plugging into a general training process and conducting the data-oriented, model-oriented and loss-oriented curricula. Furthermore, we present empirical results obtained by CurML to demonstrate the advantages of our library. The code is available online at https://github.com/THUMNLab/CurML.

Auxiliary learning is a widely adopted practice in deep learning, which aims to improve the model performance on the primary task by exploiting the beneficial information in the auxiliary loss. Existing auxiliary learning methods only focus on balancing the auxiliary loss and the primary loss, ignoring the module-level auxiliary influence, i.e., an auxiliary loss will be beneficial for optimizing specific modules within the model but harmful to others, failing to make full use of auxiliary information. To tackle the problem, we propose a Module-Aware Optimization approach for Auxiliary Learning (MAOAL). The proposed approach considers the module-level influence through the learnable module-level auxiliary importance, i.e., the importance of each auxiliary loss to each module. Specifically, the proposed approach jointly optimizes the module-level auxiliary importance and the model parameters in a bi-level manner. In the lower optimization, the model parameters are optimized with the importance parameterized gradient, while in the upper optimization, the module-level auxiliary importance is updated with the implicit gradient from a small developing dataset. Extensive experiments show that our proposed MAOAL method consistently outperforms state-of-the-art baselines for different auxiliary losses on various datasets, demonstrating that our method can serve as a powerful generic tool for auxiliary learning.

Discovering useful information formed by various hidden factors in multimodal data has been of great importance for recommender systems to improve both model performance and recommendation explainability. These factors hidden in multimodal data are highly entangled in a complex manner, posing great challenges in uncovering their entanglement during representation learning for recommendation. However, existing literature on disentangled representation learning only pays attention to unimodal data, failing to uncover the complex and entangled factors in multimodal data. In this paper, we study the problem of multimodal disentangled representation for recommendation in a weakly supervised manner, to the best of our knowledge, for the first time. We propose a multimodal disentangled recommendation (MDR) model that can learn well-disentangled representations carrying both complementary and common information from different modalities, such that both recommendation accuracy and representation explainability can be increased. Experimental results demonstrate the superiority of our MDR model in terms of both recommendation performance and explainabilty on various real-world datasets.

Learning disentangled representations for user intentions from multi-feedback (i.e., positive and negative feedback) can enhance the accuracy and explainability of recommendation algorithms. However, learning such disentangled representations from multi-feedback data is challenging because i) multi-feedback is complex: there exist complex relations among different types of feedback (e.g., click, unclick, and dislike, etc) as well as various user intentions, and ii) multi-feedback is noisy: there exists noisy (useless) information both in features and labels, which may deteriorate the recommendation performance. Existing works on disentangled representation learning only focus on positive feedback, failing to handle the complex relations and noise hidden in multi-feedback data. To solve this problem, in this work we propose a Curriculum Disentangled Recommendation (CDR) model that is capable of efficiently learning disentangled representations from complex and noisy multi-feedback for better recommendation. Concretely, we design a co-filtering dynamic routing mechanism that simultaneously captures the complex relations among different behavioral feedback and user intentions as well as denoise the representations in the feature level. We then present an adjustable self-evaluating curriculum that is able to evaluate sample difficulties for better model training and conduct denoising in the label level via disregarding useless information. Our extensive experiments on several real-world datasets demonstrate that the proposed CDR model can significantly outperform several state-of-the-art methods in terms of recommendation accuracy.