Shentong Mo

Paul Pu Liang, Yiwei Lyu, Xiang Fan et al. · cmu, uw

Many real-world problems are inherently multimodal, from spoken language, gestures, and paralinguistics humans use to communicate, to force, proprioception, and visual sensors on robots. While there has been an explosion of interest in multimodal learning, these methods are focused on a small set of modalities primarily in language, vision, and audio. In order to accelerate generalization towards diverse and understudied modalities, this paper studies efficient representation learning for high-modality scenarios involving a large set of diverse modalities. Since adding new models for every new modality becomes prohibitively expensive, a critical technical challenge is heterogeneity quantification: how can we measure which modalities encode similar information and interactions in order to permit parameter sharing with previous modalities? This paper proposes two new information theoretic metrics for heterogeneity quantification: (1) modality heterogeneity studies how similar 2 modalities {X1,X2} are by measuring how much information can be transferred from X1 to X2, while (2) interaction heterogeneity studies how similarly pairs of modalities {X1,X2}, {X3,X4} interact by measuring how much information can be transferred from fusing {X1,X2} to {X3,X4}. We show the importance of these 2 proposed metrics as a way to automatically prioritize the fusion of modalities that contain unique information or interactions. The result is a single model, HighMMT, that scales up to 10 modalities (text, image, audio, video, sensors, proprioception, speech, time-series, sets, and tables) and 15 tasks from 5 research areas. Not only does HighMMT outperform prior methods on the tradeoff between performance and efficiency, it also demonstrates a crucial scaling behavior: performance continues to improve with each modality added, and it transfers to entirely new modalities and tasks during fine-tuning.

Shentong Mo, Russ Salakhutdinov, Louis-Philippe Morency et al.

The Internet of Things (IoT) network integrating billions of smart physical devices embedded with sensors, software, and communication technologies is a critical and rapidly expanding component of our modern world. The IoT ecosystem provides a rich source of real-world modalities such as motion, thermal, geolocation, imaging, depth, sensors, and audio to recognize the states of humans and physical objects. Machine learning presents a rich opportunity to automatically process IoT data at scale, enabling efficient inference for understanding human wellbeing, controlling physical devices, and interconnecting smart cities. To realize this potential, we introduce IoT-LM, an open-source large multisensory language model tailored for the IoT ecosystem. IoT-LM is enabled by two technical contributions: the first is MultiIoT, the most expansive unified IoT dataset to date, encompassing over 1.15 million samples from 12 modalities and 8 tasks prepared for multisensory pre-training and instruction-tuning. The second is a new multisensory multitask adapter layer to condition pre-trained large language models on multisensory IoT data. Not only does IoT-LM yield substantial improvements on 8 supervised IoT classification tasks, but it also demonstrates new interactive question-answering, reasoning, and dialog capabilities conditioned on IoT sensors. We release IoT-LM's data sources and new multisensory language modeling framework.

Shentong Mo, Sukmin Yun

Recent diffusion transformers have demonstrated strong image synthesis capabilities but remain inefficient to train due to weak alignment between generative and discriminative representations. While representation alignment frameworks such as REPA improve convergence by aligning noisy denoising features with pretrained visual encoders, their externally supervised alignment loss is static and lacks adaptivity during training and inference. Existing methods rely on fixed cosine alignment or contrastive objectives, which cannot dynamically balance representation consistency and generation quality, resulting in limited discriminative benefit and failing to optimize alignment in a task-adaptive manner. To address this, we propose VRPO, a reinforcement-based optimization strategy that replaces REPA's static alignment loss with a generative representation policy optimization objective. Instead of enforcing a fixed similarity constraint, VRPO treats representation alignment as a reward-guided process: the model receives adaptive rewards based on generation fidelity, perceptual quality, and semantic coherence between the diffusion features and pretrained visual embeddings. This formulation enables the generator to continuously refine its internal representations toward semantically meaningful directions while improving image quality. Our VRPO-driven training seamlessly integrates into diffusion transformers, introducing negligible computation cost and preserving full compatibility with SiT and DiT architectures. Extensive experiments on ImageNet-256x256 demonstrate that our VRPO-Alignment substantially enhances both convergence and fidelity, achieving up to +1.8 FID improvement and 2.3x faster training compared to REPA under identical compute budgets.

Shentong Mo, Yapeng Tian

Sound source localization is a typical and challenging task that predicts the location of sound sources in a video. Previous single-source methods mainly used the audio-visual association as clues to localize sounding objects in each image. Due to the mixed property of multiple sound sources in the original space, there exist rare multi-source approaches to localizing multiple sources simultaneously, except for one recent work using a contrastive random walk in the graph with images and separated sound as nodes. Despite their promising performance, they can only handle a fixed number of sources, and they cannot learn compact class-aware representations for individual sources. To alleviate this shortcoming, in this paper, we propose a novel audio-visual grouping network, namely AVGN, that can directly learn category-wise semantic features for each source from the input audio mixture and image to localize multiple sources simultaneously. Specifically, our AVGN leverages learnable audio-visual class tokens to aggregate class-aware source features. Then, the aggregated semantic features for each source can be used as guidance to localize the corresponding visual regions. Compared to existing multi-source methods, our new framework can localize a flexible number of sources and disentangle category-aware audio-visual representations for individual sound sources. We conduct extensive experiments on MUSIC, VGGSound-Instruments, and VGG-Sound Sources benchmarks. The results demonstrate that the proposed AVGN can achieve state-of-the-art sounding object localization performance on both single-source and multi-source scenarios. Code is available at \url{https://github.com/stoneMo/AVGN}.

Fangyi Chen, Han Zhang, Zaiwang Li et al. · cmu

To make full use of computer vision technology in stores, it is required to consider the actual needs that fit the characteristics of the retail scene. Pursuing this goal, we introduce the United Retail Datasets (Unitail), a large-scale benchmark of basic visual tasks on products that challenges algorithms for detecting, reading, and matching. With 1.8M quadrilateral-shaped instances annotated, the Unitail offers a detection dataset to align product appearance better. Furthermore, it provides a gallery-style OCR dataset containing 1454 product categories, 30k text regions, and 21k transcriptions to enable robust reading on products and motivate enhanced product matching. Besides benchmarking the datasets using various state-of-the-arts, we customize a new detector for product detection and provide a simple OCR-based matching solution that verifies its effectiveness.

Weiguo Pian, Shentong Mo, Yunhui Guo et al.

In this paper, we introduce audio-visual class-incremental learning, a class-incremental learning scenario for audio-visual video recognition. We demonstrate that joint audio-visual modeling can improve class-incremental learning, but current methods fail to preserve semantic similarity between audio and visual features as incremental step grows. Furthermore, we observe that audio-visual correlations learned in previous tasks can be forgotten as incremental steps progress, leading to poor performance. To overcome these challenges, we propose AV-CIL, which incorporates Dual-Audio-Visual Similarity Constraint (D-AVSC) to maintain both instance-aware and class-aware semantic similarity between audio-visual modalities and Visual Attention Distillation (VAD) to retain previously learned audio-guided visual attentive ability. We create three audio-visual class-incremental datasets, AVE-Class-Incremental (AVE-CI), Kinetics-Sounds-Class-Incremental (K-S-CI), and VGGSound100-Class-Incremental (VS100-CI) based on the AVE, Kinetics-Sounds, and VGGSound datasets, respectively. Our experiments on AVE-CI, K-S-CI, and VS100-CI demonstrate that AV-CIL significantly outperforms existing class-incremental learning methods in audio-visual class-incremental learning. Code and data are available at: https://github.com/weiguoPian/AV-CIL_ICCV2023.

Shentong Mo, Pedro Morgado

Unsupervised audio-visual source localization aims at localizing visible sound sources in a video without relying on ground-truth localization for training. Previous works often seek high audio-visual similarities for likely positive (sounding) regions and low similarities for likely negative regions. However, accurately distinguishing between sounding and non-sounding regions is challenging without manual annotations. In this work, we propose a simple yet effective approach for Easy Visual Sound Localization, namely EZ-VSL, without relying on the construction of positive and/or negative regions during training. Instead, we align audio and visual spaces by seeking audio-visual representations that are aligned in, at least, one location of the associated image, while not matching other images, at any location. We also introduce a novel object guided localization scheme at inference time for improved precision. Our simple and effective framework achieves state-of-the-art performance on two popular benchmarks, Flickr SoundNet and VGG-Sound Source. In particular, we improve the CIoU of the Flickr SoundNet test set from 76.80% to 83.94%, and on the VGG-Sound Source dataset from 34.60% to 38.85%. The code is available at https://github.com/stoneMo/EZ-VSL.

Shentong Mo, Pedro Morgado

Audio-visual source localization is a challenging task that aims to predict the location of visual sound sources in a video. Since collecting ground-truth annotations of sounding objects can be costly, a plethora of weakly-supervised localization methods that can learn from datasets with no bounding-box annotations have been proposed in recent years, by leveraging the natural co-occurrence of audio and visual signals. Despite significant interest, popular evaluation protocols have two major flaws. First, they allow for the use of a fully annotated dataset to perform early stopping, thus significantly increasing the annotation effort required for training. Second, current evaluation metrics assume the presence of sound sources at all times. This is of course an unrealistic assumption, and thus better metrics are necessary to capture the model's performance on (negative) samples with no visible sound sources. To accomplish this, we extend the test set of popular benchmarks, Flickr SoundNet and VGG-Sound Sources, in order to include negative samples, and measure performance using metrics that balance localization accuracy and recall. Using the new protocol, we conducted an extensive evaluation of prior methods, and found that most prior works are not capable of identifying negatives and suffer from significant overfitting problems (rely heavily on early stopping for best results). We also propose a new approach for visual sound source localization that addresses both these problems. In particular, we found that, through extreme visual dropout and the use of momentum encoders, the proposed approach combats overfitting effectively, and establishes a new state-of-the-art performance on both Flickr SoundNet and VGG-Sound Source. Code and pre-trained models are available at https://github.com/stoneMo/SLAVC.

Shentong Mo, Enze Xie, Ruihang Chu et al.

Recent Diffusion Transformers (e.g., DiT) have demonstrated their powerful effectiveness in generating high-quality 2D images. However, it is still being determined whether the Transformer architecture performs equally well in 3D shape generation, as previous 3D diffusion methods mostly adopted the U-Net architecture. To bridge this gap, we propose a novel Diffusion Transformer for 3D shape generation, namely DiT-3D, which can directly operate the denoising process on voxelized point clouds using plain Transformers. Compared to existing U-Net approaches, our DiT-3D is more scalable in model size and produces much higher quality generations. Specifically, the DiT-3D adopts the design philosophy of DiT but modifies it by incorporating 3D positional and patch embeddings to adaptively aggregate input from voxelized point clouds. To reduce the computational cost of self-attention in 3D shape generation, we incorporate 3D window attention into Transformer blocks, as the increased 3D token length resulting from the additional dimension of voxels can lead to high computation. Finally, linear and devoxelization layers are used to predict the denoised point clouds. In addition, our transformer architecture supports efficient fine-tuning from 2D to 3D, where the pre-trained DiT-2D checkpoint on ImageNet can significantly improve DiT-3D on ShapeNet. Experimental results on the ShapeNet dataset demonstrate that the proposed DiT-3D achieves state-of-the-art performance in high-fidelity and diverse 3D point cloud generation. In particular, our DiT-3D decreases the 1-Nearest Neighbor Accuracy of the state-of-the-art method by 4.59 and increases the Coverage metric by 3.51 when evaluated on Chamfer Distance.

Shentong Mo, Weiguo Pian, Yapeng Tian

Continual learning is a challenging problem in which models need to be trained on non-stationary data across sequential tasks for class-incremental learning. While previous methods have focused on using either regularization or rehearsal-based frameworks to alleviate catastrophic forgetting in image classification, they are limited to a single modality and cannot learn compact class-aware cross-modal representations for continual audio-visual learning. To address this gap, we propose a novel class-incremental grouping network (CIGN) that can learn category-wise semantic features to achieve continual audio-visual learning. Our CIGN leverages learnable audio-visual class tokens and audio-visual grouping to continually aggregate class-aware features. Additionally, it utilizes class tokens distillation and continual grouping to prevent forgetting parameters learned from previous tasks, thereby improving the model's ability to capture discriminative audio-visual categories. We conduct extensive experiments on VGGSound-Instruments, VGGSound-100, and VGG-Sound Sources benchmarks. Our experimental results demonstrate that the CIGN achieves state-of-the-art audio-visual class-incremental learning performance. Code is available at https://github.com/stoneMo/CIGN.

Ruihang Chu, Enze Xie, Shentong Mo et al.

We introduce a new diffusion-based approach for shape completion on 3D range scans. Compared with prior deterministic and probabilistic methods, we strike a balance between realism, multi-modality, and high fidelity. We propose DiffComplete by casting shape completion as a generative task conditioned on the incomplete shape. Our key designs are two-fold. First, we devise a hierarchical feature aggregation mechanism to inject conditional features in a spatially-consistent manner. So, we can capture both local details and broader contexts of the conditional inputs to control the shape completion. Second, we propose an occupancy-aware fusion strategy in our model to enable the completion of multiple partial shapes and introduce higher flexibility on the input conditions. DiffComplete sets a new SOTA performance (e.g., 40% decrease on l_1 error) on two large-scale 3D shape completion benchmarks. Our completed shapes not only have a realistic outlook compared with the deterministic methods but also exhibit high similarity to the ground truths compared with the probabilistic alternatives. Further, DiffComplete has strong generalizability on objects of entirely unseen classes for both synthetic and real data, eliminating the need for model re-training in various applications.

Jiantao Wu, Shentong Mo, Muhammad Awais et al.

Self-supervised pretraining (SSP) has emerged as a popular technique in machine learning, enabling the extraction of meaningful feature representations without labelled data. In the realm of computer vision, pretrained vision transformers (ViTs) have played a pivotal role in advancing transfer learning. Nonetheless, the escalating cost of finetuning these large models has posed a challenge due to the explosion of model size. This study endeavours to evaluate the effectiveness of pure self-supervised learning (SSL) techniques in computer vision tasks, obviating the need for finetuning, with the intention of emulating human-like capabilities in generalisation and recognition of unseen objects. To this end, we propose an evaluation protocol for zero-shot segmentation based on a prompting patch. Given a point on the target object as a prompt, the algorithm calculates the similarity map between the selected patch and other patches, upon that, a simple thresholding is applied to segment the target. Another evaluation is intra-object and inter-object similarity to gauge discriminatory ability of SSP ViTs. Insights from zero-shot segmentation from prompting and discriminatory abilities of SSP led to the design of a simple SSP approach, termed MMC. This approaches combines Masked image modelling for encouraging similarity of local features, Momentum based self-distillation for transferring semantics from global to local features, and global Contrast for promoting semantics of global features, to enhance discriminative representations of SSP ViTs. Consequently, our proposed method significantly reduces the overlap of intra-object and inter-object similarities, thereby facilitating effective object segmentation within an image. Our experiments reveal that MMC delivers top-tier results in zero-shot semantic segmentation across various datasets.

Jiantao Wu, Shentong Mo, Xiang Yang et al.

One major challenge of disentanglement learning with variational autoencoders is the trade-off between disentanglement and reconstruction fidelity. Previous studies, which increase the information bottleneck during training, tend to lose the constraint of disentanglement, leading to the information diffusion problem. In this paper, we present a novel framework for disentangled representation learning, DeVAE, which utilizes hierarchical latent spaces with decreasing information bottlenecks across these spaces. The key innovation of our approach lies in connecting the hierarchical latent spaces through disentanglement-invariant transformations, allowing the sharing of disentanglement properties among spaces while maintaining an acceptable level of reconstruction performance. We demonstrate the effectiveness of DeVAE in achieving a balance between disentanglement and reconstruction through a series of experiments and ablation studies on dSprites and Shapes3D datasets. Code is available at https://github.com/erow/disentanglement_lib/tree/pytorch#devae.

Shentong Mo, Louis-Philippe Morency, Russ Salakhutdinov et al.

The next generation of machine learning systems must be adept at perceiving and interacting with the physical world through a diverse array of sensory channels. Commonly referred to as the `Internet of Things (IoT)' ecosystem, sensory data from motion, thermal, geolocation, depth, wireless signals, video, and audio are increasingly used to model the states of physical environments and the humans inside them. Despite the potential for understanding human wellbeing, controlling physical devices, and interconnecting smart cities, the community has seen limited benchmarks for building machine learning systems for IoT. Existing efforts are often specialized to a single sensory modality or prediction task, which makes it difficult to study and train large-scale models across many IoT sensors and tasks. To accelerate the development of new machine learning technologies for IoT, this paper proposes MultiIoT, the most expansive and unified IoT benchmark to date, encompassing over 1.15 million samples from 12 modalities and 8 real-world tasks. MultiIoT introduces unique challenges involving (1) generalizable learning from many sensory modalities, (2) multimodal interactions across long temporal ranges, (3) extreme heterogeneity due to unique structure and noise topologies in real-world sensors, and (4) complexity during training and inference. We evaluate a comprehensive set of models on MultiIoT, including modality and task-specific methods, multisensory and multitask supervised models, and large multisensory foundation models. Our results highlight opportunities for ML to make a significant impact in IoT, but many challenges in scalable learning from heterogeneous, long-range, and imperfect sensory modalities still persist. We release all code and data to accelerate future research in machine learning for IoT.

Shentong Mo, Jingfei Xia, Xiaoqing Tan et al.

Spatio-temporal action recognition has been a challenging task that involves detecting where and when actions occur. Current state-of-the-art action detectors are mostly anchor-based, requiring sensitive anchor designs and huge computations due to calculating large numbers of anchor boxes. Motivated by nascent anchor-free approaches, we propose Point3D, a flexible and computationally efficient network with high precision for spatio-temporal action recognition. Our Point3D consists of a Point Head for action localization and a 3D Head for action classification. Firstly, Point Head is used to track center points and knot key points of humans to localize the bounding box of an action. These location features are then piped into a time-wise attention to learn long-range dependencies across frames. The 3D Head is later deployed for the final action classification. Our Point3D achieves state-of-the-art performance on the JHMDB, UCF101-24, and AVA benchmarks in terms of frame-mAP and video-mAP. Comprehensive ablation studies also demonstrate the effectiveness of each module proposed in our Point3D.

Shentong Mo, Miao Xin

While the recently introduced Tree of Thoughts (ToT) has heralded advancements in allowing Large Language Models (LLMs) to reason through foresight and backtracking for global decision-making, it has overlooked the inherent local uncertainties in intermediate decision points or "thoughts". These local uncertainties, intrinsic to LLMs given their potential for diverse responses, remain a significant concern in the reasoning process. Addressing this pivotal gap, we introduce the Tree of Uncertain Thoughts (TouT) - a reasoning framework tailored for LLMs. Our TouT effectively leverages Monte Carlo Dropout to quantify uncertainty scores associated with LLMs' diverse local responses at these intermediate steps. By marrying this local uncertainty quantification with global search algorithms, TouT enhances the model's precision in response generation. We substantiate our approach with rigorous experiments on two demanding planning tasks: Game of 24 and Mini Crosswords. The empirical evidence underscores TouT's superiority over both ToT and chain-of-thought prompting methods.

Jiantao Wu, Shentong Mo

Self-supervised pre-training for images without labels has recently achieved promising performance in image classification. The success of transformer-based methods, ViT and MAE, draws the community's attention to the design of backbone architecture and self-supervised task. In this work, we show that current masked image encoding models learn the underlying relationship between all objects in the whole scene, instead of a single object representation. Therefore, those methods bring a lot of compute time for self-supervised pre-training. To solve this issue, we introduce a novel object selection and division strategy to drop non-object patches for learning object-wise representations by selective reconstruction with interested region masks. We refer to this method ObjMAE. Extensive experiments on four commonly-used datasets demonstrate the effectiveness of our model in reducing the compute cost by 72% while achieving competitive performance. Furthermore, we investigate the inter-object and intra-object relationship and find that the latter is crucial for self-supervised pre-training.

Shentong Mo, Bhiksha Raj

Audio-visual segmentation is a challenging task that aims to predict pixel-level masks for sound sources in a video. Previous work applied a comprehensive manually designed architecture with countless pixel-wise accurate masks as supervision. However, these pixel-level masks are expensive and not available in all cases. In this work, we aim to simplify the supervision as the instance-level annotation, i.e., weakly-supervised audio-visual segmentation. We present a novel Weakly-Supervised Audio-Visual Segmentation framework, namely WS-AVS, that can learn multi-scale audio-visual alignment with multi-scale multiple-instance contrastive learning for audio-visual segmentation. Extensive experiments on AVSBench demonstrate the effectiveness of our WS-AVS in the weakly-supervised audio-visual segmentation of single-source and multi-source scenarios.

Shentong Mo, Daizong Liu, Wei Hu

Query-based video grounding is an important yet challenging task in video understanding, which aims to localize the target segment in an untrimmed video according to a sentence query. Most previous works achieve significant progress by addressing this task in a fully-supervised manner with segment-level labels, which require high labeling cost. Although some recent efforts develop weakly-supervised methods that only need the video-level knowledge, they generally match multiple pre-defined segment proposals with query and select the best one, which lacks fine-grained frame-level details for distinguishing frames with high repeatability and similarity within the entire video. To alleviate the above limitations, we propose a self-contrastive learning framework to address the query-based video grounding task under a weakly-supervised setting. Firstly, instead of utilizing redundant segment proposals, we propose a new grounding scheme that learns frame-wise matching scores referring to the query semantic to predict the possible foreground frames by only using the video-level annotations. Secondly, since some predicted frames (i.e., boundary frames) are relatively coarse and exhibit similar appearance to their adjacent frames, we propose a coarse-to-fine contrastive learning paradigm to learn more discriminative frame-wise representations for distinguishing the false positive frames. In particular, we iteratively explore multi-scale hard negative samples that are close to positive samples in the representation space for distinguishing fine-grained frame-wise details, thus enforcing more accurate segment grounding. Extensive experiments on two challenging benchmarks demonstrate the superiority of our proposed method compared with the state-of-the-art methods.

Shentong Mo, Zhun Sun, Chao Li

Contrastive Self-supervised Learning (CSL) is a practical solution that learns meaningful visual representations from massive data in an unsupervised approach. The ordinary CSL embeds the features extracted from neural networks onto specific topological structures. During the training progress, the contrastive loss draws the different views of the same input together while pushing the embeddings from different inputs apart. One of the drawbacks of CSL is that the loss term requires a large number of negative samples to provide better mutual information bound ideally. However, increasing the number of negative samples by larger running batch size also enhances the effects of false negatives: semantically similar samples are pushed apart from the anchor, hence downgrading downstream performance. In this paper, we tackle this problem by introducing a simple but effective contrastive learning framework. The key insight is to employ siamese-style metric loss to match intra-prototype features, while increasing the distance between inter-prototype features. We conduct extensive experiments on various benchmarks where the results demonstrate the effectiveness of our method on improving the quality of visual representations. Specifically, our unsupervised pre-trained ResNet-50 with a linear probe, out-performs the fully-supervised trained version on the ImageNet-1K dataset.

Shentong Mo, Zhun Sun, Chao Li

Contrastive self-supervised learning (CSL) with a prototypical regularization has been introduced in learning meaningful representations for downstream tasks that require strong semantic information. However, to optimize CSL with a loss that performs the prototypical regularization aggressively, e.g., the ProtoNCE loss, might cause the "coagulation" of examples in the embedding space. That is, the intra-prototype diversity of samples collapses to trivial solutions for their prototype being well-separated from others. Motivated by previous works, we propose to mitigate this phenomenon by learning Prototypical representation through Alignment, Uniformity and Correlation (PAUC). Specifically, the ordinary ProtoNCE loss is revised with: (1) an alignment loss that pulls embeddings from positive prototypes together; (2) a uniformity loss that distributes the prototypical level features uniformly; (3) a correlation loss that increases the diversity and discriminability between prototypical level features. We conduct extensive experiments on various benchmarks where the results demonstrate the effectiveness of our method in improving the quality of prototypical contrastive representations. Particularly, in the classification down-stream tasks with linear probes, our proposed method outperforms the state-of-the-art instance-wise and prototypical contrastive learning methods on the ImageNet-100 dataset by 2.96% and the ImageNet-1K dataset by 2.46% under the same settings of batch size and epochs.

Shentong Mo, Pedro Morgado

Audio-visual generalized zero-shot learning is a rapidly advancing domain that seeks to understand the intricate relations between audio and visual cues within videos. The overarching goal is to leverage insights from seen classes to identify instances from previously unseen ones. Prior approaches primarily utilized synchronized auto-encoders to reconstruct audio-visual attributes, which were informed by cross-attention transformers and projected text embeddings. However, these methods fell short of effectively capturing the intricate relationship between cross-modal features and class-label embeddings inherent in pre-trained language-aligned embeddings. To circumvent these bottlenecks, we introduce a simple yet effective framework for Easy Audio-Visual Generalized Zero-shot Learning, named EZ-AVGZL, that aligns audio-visual embeddings with transformed text representations. It utilizes a single supervised text audio-visual contrastive loss to learn an alignment between audio-visual and textual modalities, moving away from the conventional approach of reconstructing cross-modal features and text embeddings. Our key insight is that while class name embeddings are well aligned with language-based audio-visual features, they don't provide sufficient class separation to be useful for zero-shot learning. To address this, our method leverages differential optimization to transform class embeddings into a more discriminative space while preserving the semantic structure of language representations. We conduct extensive experiments on VGGSound-GZSL, UCF-GZSL, and ActivityNet-GZSL benchmarks. Our results demonstrate that our EZ-AVGZL achieves state-of-the-art performance in audio-visual generalized zero-shot learning.

Shentong Mo, Paul Pu Liang

Biomedical data is inherently multimodal, consisting of electronic health records, medical imaging, digital pathology, genome sequencing, wearable sensors, and more. The application of artificial intelligence tools to these multifaceted sensing technologies has the potential to revolutionize the prognosis, diagnosis, and management of human health and disease. However, current approaches to biomedical AI typically only train and evaluate with one or a small set of medical modalities and tasks. This limitation hampers the development of comprehensive tools that can leverage the rich interconnected information across many heterogeneous biomedical sensors. To address this challenge, we present MultiMed, a benchmark designed to evaluate and enable large-scale learning across a wide spectrum of medical modalities and tasks. MultiMed consists of 2.56 million samples across ten medical modalities such as medical reports, pathology, genomics, and protein data, and is structured into eleven challenging tasks, including disease prognosis, protein structure prediction, and medical question answering. Using MultiMed, we conduct comprehensive experiments benchmarking state-of-the-art unimodal, multimodal, and multitask models. Our analysis highlights the advantages of training large-scale medical models across many related modalities and tasks. Moreover, MultiMed enables studies of generalization across related medical concepts, robustness to real-world noisy data and distribution shifts, and novel modality combinations to improve prediction performance. MultiMed will be publicly available and regularly updated and welcomes inputs from the community.

Shentong Mo, Yapeng Tian

Recently, audio-visual separation approaches have taken advantage of the natural synchronization between the two modalities to boost audio source separation performance. They extracted high-level semantics from visual inputs as the guidance to help disentangle sound representation for individual sources. Can we directly learn to disentangle the individual semantics from the sound itself? The dilemma is that multiple sound sources are mixed together in the original space. To tackle the difficulty, in this paper, we present a novel Semantic Grouping Network, termed as SGN, that can directly disentangle sound representations and extract high-level semantic information for each source from input audio mixture. Specifically, SGN aggregates category-wise source features through learnable class tokens of sounds. Then, the aggregated semantic features can be used as the guidance to separate the corresponding audio sources from the mixture. We conducted extensive experiments on music-only and universal sound separation benchmarks: MUSIC, FUSS, MUSDB18, and VGG-Sound. The results demonstrate that our SGN significantly outperforms previous audio-only methods and audio-visual models without utilizing additional visual cues.

Shentong Mo, Haofan Wang

Visual sound localization is a typical and challenging problem that predicts the location of objects corresponding to the sound source in a video. Previous methods mainly used the audio-visual association between global audio and one-scale visual features to localize sounding objects in each image. Despite their promising performance, they omitted multi-scale visual features of the corresponding image, and they cannot learn discriminative regions compared to ground truths. To address this issue, we propose a novel multi-scale multi-instance visual sound localization framework, namely M2VSL, that can directly learn multi-scale semantic features associated with sound sources from the input image to localize sounding objects. Specifically, our M2VSL leverages learnable multi-scale visual features to align audio-visual representations at multi-level locations of the corresponding image. We also introduce a novel multi-scale multi-instance transformer to dynamically aggregate multi-scale cross-modal representations for visual sound localization. We conduct extensive experiments on VGGSound-Instruments, VGG-Sound Sources, and AVSBench benchmarks. The results demonstrate that the proposed M2VSL can achieve state-of-the-art performance on sounding object localization and segmentation.

Shentong Mo, Zhun Sun, Chao Li

Data augmentation has become a standard component of vision pre-trained models to capture the invariance between augmented views. In practice, augmentation techniques that mask regions of a sample with zero/mean values or patches from other samples are commonly employed in pre-trained models with self-/semi-/fully-supervised contrastive losses. However, the underlying mechanism behind the effectiveness of these augmentation techniques remains poorly explored. To investigate the problems, we conduct an empirical study to quantify how data augmentation affects performance. Concretely, we apply 4 types of data augmentations termed with Random Erasing, CutOut, CutMix and MixUp to a series of self-/semi-/fully- supervised pre-trained models. We report their performance on vision tasks such as image classification, object detection, instance segmentation, and semantic segmentation. We then explicitly evaluate the invariance and diversity of the feature embedding. We observe that: 1) Masking regions of the images decreases the invariance of the learned feature embedding while providing a more considerable diversity. 2) Manual annotations do not change the invariance or diversity of the learned feature embedding. 3) The MixUp approach improves the diversity significantly, with only a marginal decrease in terms of the invariance.

Shentong Mo, Jingfei Xia, Ihor Markevych

Visual and linguistic pre-training aims to learn vision and language representations together, which can be transferred to visual-linguistic downstream tasks. However, there exists semantic confusion between language and vision during the pre-training stage. Moreover, current pre-trained models tend to take lots of computation resources for fine-tuning when transferred to downstream tasks. In this work, we present a simple but effective approach for learning Contrastive and Adaptive representations of Vision and Language, namely CAVL. Specifically, we introduce a pair-wise contrastive loss to learn alignments between the whole sentence and each image in the same batch during the pre-training process. At the fine-tuning stage, we introduce two lightweight adaptation networks to reduce model parameters and increase training speed for saving computation resources. We evaluate our CAVL on six main downstream tasks, including Visual Question Answering (VQA), Visual Commonsense Reasoning (VCR), Natural Language for Visual Reasoning (NLVR), Region-to-Phrase Grounding (RPG), Text-to-Image Retrieval (TIR), and Zero-shot Text-to-Image Retrieval (ZS-TIR). Compared to baselines, we achieve superior performance and reduce the fine-tuning time by a large margin (in particular, 76.17%). Extensive experiments and ablation studies demonstrate the efficiency of contrastive pre-training and adaptive fine-tuning proposed in our CAVL.

Shentong Mo, Xufang Luo, Dongsheng Li

Parameter-efficient fine-tuning has demonstrated promising results across various visual adaptation tasks, such as classification and segmentation. Typically, prompt tuning techniques have harnessed knowledge from a single pre-trained model, whether from a general or a specialized medical domain. However, this approach typically overlooks the potential synergies that could arise from integrating diverse domain knowledge within the same tuning process. In this work, we propose a novel Mixture-of-Experts prompt tuning method called pMoE, which leverages the strengths of multiple expert domains through expert-specialized prompt tokens and the learnable dispatcher, effectively combining their expertise in a unified model framework. Our pMoE introduces expert-specific prompt tokens and utilizes a dynamic token dispatching mechanism at various prompt layers to optimize the contribution of each domain expert during the adaptation phase. By incorporating both domain knowledge from diverse experts, the proposed pMoE significantly enhances the model's versatility and applicability to a broad spectrum of tasks. We conduct extensive experiments across 47 adaptation tasks, including both classification and segmentation in general and medical domains. The results demonstrate that our pMoE not only achieves superior performance with a large margin of improvements but also offers an optimal trade-off between computational efficiency and adaptation effectiveness compared to existing methods.

Shentong Mo, Lanqing Li

Generative models for de novo protein backbone design have achieved remarkable success in creating novel protein structures. However, these diffusion-based approaches remain computationally intensive and slower than desired for large-scale structural exploration. While recent efforts like Proteina have introduced flow-matching to improve sampling efficiency, the potential of tokenization for structural compression and acceleration remains largely unexplored in the protein domain. In this work, we present SaDiT, a novel framework that accelerates protein backbone generation by integrating SaProt Tokenization with a Diffusion Transformer (DiT) architecture. SaDiT leverages a discrete latent space to represent protein geometry, significantly reducing the complexity of the generation process while maintaining theoretical SE(3) equivalence. To further enhance efficiency, we introduce an IPA Token Cache mechanism that optimizes the Invariant Point Attention (IPA) layers by reusing computed token states during iterative sampling. Experimental results demonstrate that SaDiT outperforms state-of-the-art models, including RFDiffusion and Proteina, in both computational speed and structural viability. We evaluate our model across unconditional backbone generation and fold-class conditional generation tasks, where SaDiT shows superior ability to capture complex topological features with high designability.

Jiantao Wu, Shentong Mo, Sara Atito et al.

Recently, masked image modeling (MIM), an important self-supervised learning (SSL) method, has drawn attention for its effectiveness in learning data representation from unlabeled data. Numerous studies underscore the advantages of MIM, highlighting how models pretrained on extensive datasets can enhance the performance of downstream tasks. However, the high computational demands of pretraining pose significant challenges, particularly within academic environments, thereby impeding the SSL research progress. In this study, we propose efficient training recipes for MIM based SSL that focuses on mitigating data loading bottlenecks and employing progressive training techniques and other tricks to closely maintain pretraining performance. Our library enables the training of a MAE-Base/16 model on the ImageNet 1K dataset for 800 epochs within just 18 hours, using a single machine equipped with 8 A100 GPUs. By achieving speed gains of up to 5.8 times, this work not only demonstrates the feasibility of conducting high-efficiency SSL training but also paves the way for broader accessibility and promotes advancement in SSL research particularly for prototyping and initial testing of SSL ideas. The code is available in https://github.com/erow/FastSSL.

Shentong Mo, Sukmin Yun

Unified multimodal pretraining has emerged as a promising paradigm for jointly modeling language and vision within a single foundation model. However, existing approaches largely rely on implicit or indirect alignment signals and remain suboptimal for simultaneously supporting multimodal understanding and generation, particularly in settings that require fine-grained language-visual reasoning and controllable generation. In this work, we propose LVRPO, a language-visual reinforcement-based preference optimization framework that explicitly aligns language and visual representations using Group Relative Policy Optimization (GRPO). Instead of introducing additional alignment losses at the representation level, LVRPO directly optimizes multimodal model behaviors through preference-driven reinforcement signals, encouraging consistent and semantically grounded interactions between language and vision across both understanding and generation tasks. This formulation enables effective alignment without requiring auxiliary encoders or handcrafted cross-modal objectives, and naturally extends to diverse multimodal capabilities. Empirically, LVRPO consistently outperforms strong unified-pretraining baselines on a broad suite of benchmarks spanning multimodal understanding, generation, and reasoning.

Weiguo Pian, Yiyang Nan, Shijian Deng et al.

In this paper, we introduce a novel continual audio-visual sound separation task, aiming to continuously separate sound sources for new classes while preserving performance on previously learned classes, with the aid of visual guidance. This problem is crucial for practical visually guided auditory perception as it can significantly enhance the adaptability and robustness of audio-visual sound separation models, making them more applicable for real-world scenarios where encountering new sound sources is commonplace. The task is inherently challenging as our models must not only effectively utilize information from both modalities in current tasks but also preserve their cross-modal association in old tasks to mitigate catastrophic forgetting during audio-visual continual learning. To address these challenges, we propose a novel approach named ContAV-Sep (\textbf{Cont}inual \textbf{A}udio-\textbf{V}isual Sound \textbf{Sep}aration). ContAV-Sep presents a novel Cross-modal Similarity Distillation Constraint (CrossSDC) to uphold the cross-modal semantic similarity through incremental tasks and retain previously acquired knowledge of semantic similarity in old models, mitigating the risk of catastrophic forgetting. The CrossSDC can seamlessly integrate into the training process of different audio-visual sound separation frameworks. Experiments demonstrate that ContAV-Sep can effectively mitigate catastrophic forgetting and achieve significantly better performance compared to other continual learning baselines for audio-visual sound separation. Code is available at: \url{https://github.com/weiguoPian/ContAV-Sep_NeurIPS2024}.

Shentong Mo, Zehua Chen, Jun Zhu

Recent advances in video-audio (V-A) understanding and generation have increasingly relied on joint V-A embeddings, which serve as the foundation for tasks such as cross-modal retrieval and generation. While prior methods like CAVP effectively model semantic and temporal correspondences between modalities using contrastive objectives, their performance remains suboptimal. A key limitation is the insufficient modeling of the dense, multi-scale nature of both video and audio signals, correspondences often span fine- to coarse-grained spatial-temporal structures, which are underutilized in existing frameworks. To this end, we propose GMS-CAVP, a novel framework that combines Multi-Scale Video-Audio Alignment and Multi-Scale Spatial-Temporal Diffusion-based pretraining objectives to enhance V-A correspondence modeling. First, GMS-CAVP introduces a multi-scale contrastive learning strategy that captures semantic and temporal relations across varying granularities. Second, we go beyond traditional contrastive learning by incorporating a diffusion-based generative objective, enabling modality translation and synthesis between video and audio. This unified discriminative-generative formulation facilitates deeper cross-modal understanding and paves the way for high-fidelity generation. Extensive experiments on VGGSound, AudioSet, and Panda70M demonstrate that GMS-CAVP outperforms previous methods in generation and retrieval.

Xiaokang Chen, Mingyu Ding, Xiaodi Wang et al.

We present a novel masked image modeling (MIM) approach, context autoencoder (CAE), for self-supervised representation pretraining. We pretrain an encoder by making predictions in the encoded representation space. The pretraining tasks include two tasks: masked representation prediction - predict the representations for the masked patches, and masked patch reconstruction - reconstruct the masked patches. The network is an encoder-regressor-decoder architecture: the encoder takes the visible patches as input; the regressor predicts the representations of the masked patches, which are expected to be aligned with the representations computed from the encoder, using the representations of visible patches and the positions of visible and masked patches; the decoder reconstructs the masked patches from the predicted encoded representations. The CAE design encourages the separation of learning the encoder (representation) from completing the pertaining tasks: masked representation prediction and masked patch reconstruction tasks, and making predictions in the encoded representation space empirically shows the benefit to representation learning. We demonstrate the effectiveness of our CAE through superior transfer performance in downstream tasks: semantic segmentation, object detection and instance segmentation, and classification. The code will be available at https://github.com/Atten4Vis/CAE.

Shentong Mo, Xiaoqing Tan, Jingfei Xia et al.

Spatiotemporal action recognition deals with locating and classifying actions in videos. Motivated by the latest state-of-the-art real-time object detector You Only Watch Once (YOWO), we aim to modify its structure to increase action detection precision and reduce computational time. Specifically, we propose four novel approaches in attempts to improve YOWO and address the imbalanced class issue in videos by modifying the loss function. We consider two moderate-sized datasets to apply our modification of YOWO - the popular Joint-annotated Human Motion Data Base (J-HMDB-21) and a private dataset of restaurant video footage provided by a Carnegie Mellon University-based startup, Agot.AI. The latter involves fast-moving actions with small objects as well as unbalanced data classes, making the task of action localization more challenging. We implement our proposed methods in the GitHub repository https://github.com/stoneMo/YOWOv2.

Shentong Mo

Designing effective reward functions is a cornerstone of reinforcement learning (RL), yet it remains a challenging and labor-intensive process due to the inefficiencies and inconsistencies inherent in traditional methods. Existing methods often rely on extensive manual design and evaluation steps, which are prone to redundancy and overlook local uncertainties at intermediate decision points. To address these challenges, we propose the Chain of Uncertain Rewards (CoUR), a novel framework that integrates large language models (LLMs) to streamline reward function design and evaluation in RL environments. Specifically, our CoUR introduces code uncertainty quantification with a similarity selection mechanism that combines textual and semantic analyses to identify and reuse the most relevant reward function components. By reducing redundant evaluations and leveraging Bayesian optimization on decoupled reward terms, CoUR enables a more efficient and robust search for optimal reward feedback. We comprehensively evaluate CoUR across nine original environments from IsaacGym and all 20 tasks from the Bidexterous Manipulation benchmark. The experimental results demonstrate that CoUR not only achieves better performance but also significantly lowers the cost of reward evaluations.

Shentong Mo, Sukmin Yun

Generative models have made it possible to synthesize highly realistic images, potentially providing an abundant data source for training machine learning models. Despite the advantages of these synthesizable data sources, the indiscriminate use of generated images as real images for training can even cause mode collapse due to modality discrepancies between real and synthetic domains. In this paper, we propose a novel framework for discriminative use of generated images, coined GMAIL, that explicitly treats generated images as a separate modality from real images. Instead of indiscriminately replacing real images with generated ones in the pixel space, our approach bridges the two distinct modalities in the same latent space through a multi-modal learning approach. To be specific, we first fine-tune a model exclusively on generated images using a cross-modality alignment loss and then employ this aligned model to further train various vision-language models with generated images. By aligning the two modalities, our approach effectively leverages the benefits of recent advances in generative models, thereby boosting the effectiveness of generated image learning across a range of vision-language tasks. Our framework can be easily incorporated with various vision-language models, and we demonstrate its efficacy throughout extensive experiments. For example, our framework significantly improves performance on image captioning, zero-shot image retrieval, zero-shot image classification, and long caption retrieval tasks. It also shows positive generated data scaling trends and notable enhancements in the captioning performance of the large multimodal model, LLaVA.

Shentong Mo, Jing Shi, Yapeng Tian

In recent times, the focus on text-to-audio (TTA) generation has intensified, as researchers strive to synthesize audio from textual descriptions. However, most existing methods, though leveraging latent diffusion models to learn the correlation between audio and text embeddings, fall short when it comes to maintaining a seamless synchronization between the produced audio and its video. This often results in discernible audio-visual mismatches. To bridge this gap, we introduce a groundbreaking benchmark for Text-to-Audio generation that aligns with Videos, named T2AV-Bench. This benchmark distinguishes itself with three novel metrics dedicated to evaluating visual alignment and temporal consistency. To complement this, we also present a simple yet effective video-aligned TTA generation model, namely T2AV. Moving beyond traditional methods, T2AV refines the latent diffusion approach by integrating visual-aligned text embeddings as its conditional foundation. It employs a temporal multi-head attention transformer to extract and understand temporal nuances from video data, a feat amplified by our Audio-Visual ControlNet that adeptly merges temporal visual representations with text embeddings. Further enhancing this integration, we weave in a contrastive learning objective, designed to ensure that the visual-aligned text embeddings resonate closely with the audio features. Extensive evaluations on the AudioCaps and T2AV-Bench demonstrate that our T2AV sets a new standard for video-aligned TTA generation in ensuring visual alignment and temporal consistency.

Lin Zhang, Shentong Mo, Yijing Zhang et al.

Current visual generation methods can produce high quality videos guided by texts. However, effectively controlling object dynamics remains a challenge. This work explores audio as a cue to generate temporally synchronized image animations. We introduce Audio Synchronized Visual Animation (ASVA), a task animating a static image to demonstrate motion dynamics, temporally guided by audio clips across multiple classes. To this end, we present AVSync15, a dataset curated from VGGSound with videos featuring synchronized audio visual events across 15 categories. We also present a diffusion model, AVSyncD, capable of generating dynamic animations guided by audios. Extensive evaluations validate AVSync15 as a reliable benchmark for synchronized generation and demonstrate our models superior performance. We further explore AVSyncDs potential in a variety of audio synchronized generation tasks, from generating full videos without a base image to controlling object motions with various sounds. We hope our established benchmark can open new avenues for controllable visual generation. More videos on project webpage https://lzhangbj.github.io/projects/asva/asva.html.

Shentong Mo, Yapeng Tian

In recent developments, the Mamba architecture, known for its selective state space approach, has shown potential in the efficient modeling of long sequences. However, its application in image generation remains underexplored. Traditional diffusion transformers (DiT), which utilize self-attention blocks, are effective but their computational complexity scales quadratically with the input length, limiting their use for high-resolution images. To address this challenge, we introduce a novel diffusion architecture, Diffusion Mamba (DiM), which foregoes traditional attention mechanisms in favor of a scalable alternative. By harnessing the inherent efficiency of the Mamba architecture, DiM achieves rapid inference times and reduced computational load, maintaining linear complexity with respect to sequence length. Our architecture not only scales effectively but also outperforms existing diffusion transformers in both image and video generation tasks. The results affirm the scalability and efficiency of DiM, establishing a new benchmark for image and video generation techniques. This work advances the field of generative models and paves the way for further applications of scalable architectures.

Shentong Mo, Enze Xie, Yue Wu et al.

Diffusion Transformers have recently shown remarkable effectiveness in generating high-quality 3D point clouds. However, training voxel-based diffusion models for high-resolution 3D voxels remains prohibitively expensive due to the cubic complexity of attention operators, which arises from the additional dimension of voxels. Motivated by the inherent redundancy of 3D compared to 2D, we propose FastDiT-3D, a novel masked diffusion transformer tailored for efficient 3D point cloud generation, which greatly reduces training costs. Specifically, we draw inspiration from masked autoencoders to dynamically operate the denoising process on masked voxelized point clouds. We also propose a novel voxel-aware masking strategy to adaptively aggregate background/foreground information from voxelized point clouds. Our method achieves state-of-the-art performance with an extreme masking ratio of nearly 99%. Moreover, to improve multi-category 3D generation, we introduce Mixture-of-Expert (MoE) in 3D diffusion model. Each category can learn a distinct diffusion path with different experts, relieving gradient conflict. Experimental results on the ShapeNet dataset demonstrate that our method achieves state-of-the-art high-fidelity and diverse 3D point cloud generation performance. Our FastDiT-3D improves 1-Nearest Neighbor Accuracy and Coverage metrics when generating 128-resolution voxel point clouds, using only 6.5% of the original training cost.

Jiantao Wu, Sara Atito, Zhenhua Feng et al.

The training methods in AI do involve semantically distinct pairs of samples. However, their role typically is to enhance the between class separability. The actual notion of similarity is normally learned from semantically identical pairs. This paper presents SimLAP: a simple framework for learning visual representation from arbitrary pairs. SimLAP explores the possibility of learning similarity from semantically distinct sample pairs. The approach is motivated by the observation that for any pair of classes there exists a subspace in which semantically distinct samples exhibit similarity. This phenomenon can be exploited for a novel method of learning, which optimises the similarity of an arbitrary pair of samples, while simultaneously learning the enabling subspace. The feasibility of the approach will be demonstrated experimentally and its merits discussed.

Shentong Mo, Shengbang Tong

In recent advancements in unsupervised visual representation learning, the Joint-Embedding Predictive Architecture (JEPA) has emerged as a significant method for extracting visual features from unlabeled imagery through an innovative masking strategy. Despite its success, two primary limitations have been identified: the inefficacy of Exponential Moving Average (EMA) from I-JEPA in preventing entire collapse and the inadequacy of I-JEPA prediction in accurately learning the mean of patch representations. Addressing these challenges, this study introduces a novel framework, namely C-JEPA (Contrastive-JEPA), which integrates the Image-based Joint-Embedding Predictive Architecture with the Variance-Invariance-Covariance Regularization (VICReg) strategy. This integration is designed to effectively learn the variance/covariance for preventing entire collapse and ensuring invariance in the mean of augmented views, thereby overcoming the identified limitations. Through empirical and theoretical evaluations, our work demonstrates that C-JEPA significantly enhances the stability and quality of visual representation learning. When pre-trained on the ImageNet-1K dataset, C-JEPA exhibits rapid and improved convergence in both linear probing and fine-tuning performance metrics.

Shentong Mo, Yansen Wang, Xufang Luo et al. · cmu, tsinghua

Visual Prompt Tuning (VPT) techniques have gained prominence for their capacity to adapt pre-trained Vision Transformers (ViTs) to downstream visual tasks using specialized learnable tokens termed as prompts. Contemporary VPT methodologies, especially when employed with self-supervised vision transformers, often default to the introduction of new learnable prompts or gated prompt tokens predominantly sourced from the model's previous block. A pivotal oversight in such approaches is their failure to harness the potential of long-range previous blocks as sources of prompts within each self-supervised ViT. To bridge this crucial gap, we introduce Long-term Spatial Prompt Tuning (LSPT) - a revolutionary approach to visual representation learning. Drawing inspiration from the intricacies of the human brain, LSPT ingeniously incorporates long-term gated prompts. This feature serves as temporal coding, curbing the risk of forgetting parameters acquired from earlier blocks. Further enhancing its prowess, LSPT brings into play patch tokens, serving as spatial coding. This is strategically designed to perpetually amass class-conscious features, thereby fortifying the model's prowess in distinguishing and identifying visual categories. To validate the efficacy of our proposed method, we engaged in rigorous experimentation across 5 FGVC and 19 VTAB-1K benchmarks. Our empirical findings underscore the superiority of LSPT, showcasing its ability to set new benchmarks in visual prompt tuning performance.

Shentong Mo, Zehua Chen, Fan Bao et al.

Recent works in cross-modal understanding and generation, notably through models like CLAP (Contrastive Language-Audio Pretraining) and CAVP (Contrastive Audio-Visual Pretraining), have significantly enhanced the alignment of text, video, and audio embeddings via a single contrastive loss. However, these methods often overlook the bidirectional interactions and inherent noises present in each modality, which can crucially impact the quality and efficacy of cross-modal integration. To address this limitation, we introduce DiffGAP, a novel approach incorporating a lightweight generative module within the contrastive space. Specifically, our DiffGAP employs a bidirectional diffusion process tailored to bridge the cross-modal gap more effectively. This involves a denoising process on text and video embeddings conditioned on audio embeddings and vice versa, thus facilitating a more nuanced and robust cross-modal interaction. Our experimental results on VGGSound and AudioCaps datasets demonstrate that DiffGAP significantly improves performance in video/text-audio generation and retrieval tasks, confirming its effectiveness in enhancing cross-modal understanding and generation capabilities.

Shentong Mo, Xufang Luo, Yansen Wang et al. · cmu, tsinghua

Visual task adaptation has been demonstrated to be effective in adapting pre-trained Vision Transformers (ViTs) to general downstream visual tasks using specialized learnable layers or tokens. However, there is yet a large-scale benchmark to fully explore the effect of visual task adaptation on the realistic and important medical domain, particularly across diverse medical visual modalities, such as color images, X-ray, and CT. To close this gap, we present Med-VTAB, a large-scale Medical Visual Task Adaptation Benchmark consisting of 1.68 million medical images for diverse organs, modalities, and adaptation approaches. Based on Med-VTAB, we explore the scaling law of medical prompt tuning concerning tunable parameters and the generalizability of medical visual adaptation using non-medical/medical pre-train weights. Besides, we study the impact of patient ID out-of-distribution on medical visual adaptation, which is a real and challenging scenario. Furthermore, results from Med-VTAB indicate that a single pre-trained model falls short in medical task adaptation. Therefore, we introduce GMoE-Adapter, a novel method that combines medical and general pre-training weights through a gated mixture-of-experts adapter, achieving state-of-the-art results in medical visual task adaptation.

Shentong Mo, Yibing Song

Visual content and accompanied audio signals naturally formulate a joint representation to improve audio-visual (AV) related applications. While studies develop various AV representation learning frameworks, the importance of AV data alignment is usually undermined for achieving high-quality representation. We observe that an audio signal may contain background noise interference. Also, non-synchronization may appear between audio and video streams. These non-strict data alignment limits representation quality and downgrade application performance. In this paper, we propose to improve AV joint representations from a data-centric perspective by aligning audio signals to visual data. Our alignment is conducted in an agentic workflow controlled by an LLM-based assistant named AVAgent. For each input AV data pair, our AVAgent uses a multi-modal LLM to convert audio and visual data into language descriptions separately (i.e., tool use). Then, AVAgent reasons whether this paired data is aligned well and plans to edit the audio signal if needed (i.e., planning). The audio editing is executed by predefined actions that filter noise or augment data. Moreover, we use a VLM to evaluate how modified audio signals match the visual content and provide feedback to AVAgent (i.e., reflection). The tool use, planning, and reflection steps operate cyclically to become an agentic workflow where audio signals are gradually aligned to visual content. To this end, existing methods can directly leverage the aligned AV data via our agentic workflow to improve AV joint representations. The experimental results comprehensively demonstrate the state-of-the-art performance of the proposed approach against previous baselines in diverse downstream tasks.

Shentong Mo, Haofan Wang, Huaxia Li et al.

Video-language pre-training is a typical and challenging problem that aims at learning visual and textual representations from large-scale data in a self-supervised way. Existing pre-training approaches either captured the correspondence of image-text pairs or utilized temporal ordering of frames. However, they do not explicitly explore the natural synchronization between audio and the other two modalities. In this work, we propose an enhanced framework for Video-Language pre-training with Synchronized Audio, termed as VLSA, that can learn tri-modal representations in a unified self-supervised transformer. Specifically, our VLSA jointly aggregates embeddings of local patches and global tokens for video, text, and audio. Furthermore, we utilize local-patch masked modeling to learn modality-aware features, and leverage global audio matching to capture audio-guided features for video and text. We conduct extensive experiments on retrieval across text, video, and audio. Our simple model pre-trained on only 0.9M data achieves improving results against state-of-the-art baselines. In addition, qualitative visualizations vividly showcase the superiority of our VLSA in learning discriminative visual-textual representations.

Shentong Mo, Yibing Song

Coordinated audio generation based on video inputs typically requires a strict audio-visual (AV) alignment, where both semantics and rhythmics of the generated audio segments shall correspond to those in the video frames. Previous studies leverage a two-stage design where the AV encoders are firstly aligned via contrastive learning, then the encoded video representations guide the audio generation process. We observe that both contrastive learning and global video guidance are effective in aligning overall AV semantics while limiting temporally rhythmic synchronization. In this work, we propose FoleyFlow to first align unimodal AV encoders via masked modeling training, where the masked audio segments are recovered under the guidance of the corresponding video segments. After training, the AV encoders which are separately pretrained using only unimodal data are aligned with semantic and rhythmic consistency. Then, we develop a dynamic conditional flow for the final audio generation. Built upon the efficient velocity flow generation framework, our dynamic conditional flow utilizes temporally varying video features as the dynamic condition to guide corresponding audio segment generations. To this end, we extract coherent semantic and rhythmic representations during masked AV alignment, and use this representation of video segments to guide audio generation temporally. Our audio results are evaluated on the standard benchmarks and largely surpass existing results under several metrics. The superior performance indicates that FoleyFlow is effective in generating coordinated audios that are both semantically and rhythmically coherent to various video sequences.

Shentong Mo

Masked autoencoders (MAE) have recently succeeded in self-supervised vision representation learning. Previous work mainly applied custom-designed (e.g., random, block-wise) masking or teacher (e.g., CLIP)-guided masking and targets. However, they ignore the potential role of the self-training (student) model in giving feedback to the teacher for masking and targets. In this work, we present to integrate Collaborative Masking and Targets for boosting Masked AutoEncoders, namely CMT-MAE. Specifically, CMT-MAE leverages a simple collaborative masking mechanism through linear aggregation across attentions from both teacher and student models. We further propose using the output features from those two models as the collaborative target of the decoder. Our simple and effective framework pre-trained on ImageNet-1K achieves state-of-the-art linear probing and fine-tuning performance. In particular, using ViT-base, we improve the fine-tuning results of the vanilla MAE from 83.6% to 85.7%.