This paper introduces STAR-1, a high-quality, just-1k-scale safety dataset specifically designed for large reasoning models (LRMs) like DeepSeek-R1. Built on three core principles --- diversity, deliberative reasoning, and rigorous filtering --- STAR-1 aims to address the critical needs for safety alignment in LRMs. Specifically, we begin by integrating existing open-source safety datasets from diverse sources. Then, we curate safety policies to generate policy-grounded deliberative reasoning samples. Lastly, we apply a GPT-4o-based safety scoring system to select training examples aligned with best practices. Experimental results show that fine-tuning LRMs with STAR-1 leads to an average 40% improvement in safety performance across four benchmarks, while only incurring a marginal decrease (e.g., an average of 1.1%) in reasoning ability measured across five reasoning tasks. Extensive ablation studies further validate the importance of our design principles in constructing STAR-1 and analyze its efficacy across both LRMs and traditional LLMs.

Large Vision Models (LVMs) have demonstrated impressive capabilities by leveraging large-scale generative pre-training on visual sequences. However, training these models on massive datasets of single images and image sequences can be computationally expensive and limit their accessibility to researchers without substantial computational resources. This paper introduces LVM-Lite, a new two-stage learning pipeline for more efficient and effective LVM training. In the first stage, the model is pre-trained on a large corpus of single images. Subsequently, in the second stage, the model is fine-tuned on curated long image/video sequences. This decoupled training approach substantially accelerates the training process, achieving up to 2.7× speedup compared to baseline LVM training. Extensive experiments demonstrate that LVM-Lite achieves competitive performance on various generative and discriminative benchmarks while maintaining high training efficiency and strong scalability. https://github.com/UCSC-VLAA/LVM-Lite

Similar to Vision Transformers, this paper identifies artifacts also present within the feature maps of Vision Mamba. These artifacts, corresponding to high-norm tokens emerging in low-information background areas of images, appear much more severe in Vision Mamba—they exist prevalently even with the tiny-sized model and activate extensively across background regions. To mitigate this issue, we follow the prior solution of introducing register tokens into Vision Mamba. To better cope with Mamba blocks’ uni-directional inference paradigm, two key modifications are introduced: 1) evenly inserting registers throughout the input token sequence, and 2) recycling registers for final decision predictions. We term this new architecture Mamba®. Qualitative observations suggest, compared to vanilla Vision Mamba, Mamba®’s feature maps appear cleaner and more focused on semantically meaningful regions. Quantitatively, Mamba®attains stronger performance and scales better. For example, on the ImageNet benchmark, our Mamba®-B attains 83.0% accuracy, significantly outperforming Vim-B’s 81.8%; furthermore, we provide the first successful scaling to the large model size with 341M parameters, attaining competitive accuracies of 83.6% and 84.5% for 224×224 and 384×384 inputs, respectively. Additional validation on the downstream semantic segmentation task also supports Mamba®’s efficacy. Code is available at https://github.com/wangf3014/Mamba-Reg.

series models where we treat images as sequences of patch tokens and employ uni-directional language models to learn visual representations. This modeling paradigm allows us to process images in a recurrent formulation with linear complexity relative to the sequence length, which can effectively address the memory and computation explosion issues posed by high-resolution and fine-grained images. In detail, we introduce two simple designs that seamlessly integrate image inputs into the causal inference framework: a global pooling token placed at the beginning of the sequence and a flipping operation between every two layers. Extensive empirical studies highlight that compared with the existing plain architectures such as DeiT [46] and Vim [57], Adventurer offers an optimal efficiency-accuracy trade-off. For example, our Adventurer-Base attains a competitive test accuracy of 84.3% on the standard ImageNet-1k benchmark with 216 images/s training throughput, which is 3.8× and 6.2× faster than Vim and DeiT to achieve the same result. As Adventurer offers great computation and memory efficiency and allows scaling with linear complexity, we hope this architecture can benefit future explorations in modeling long sequences for high-resolution or fine-grained images. Code is available at https://github.com/wangf3014/Adventurer.

Recent advancements in large generative models, particularly diffusion-based methods, have significantly enhanced the capabilities of image editing. However, achieving precise control over image composition tasks remains a challenge. Layered representations, which allow for independent editing of image components, are essential for user-driven content creation, yet existing approaches often struggle to decompose an image into plausible layers with accurately retained transparent visual effects such as shadows and reflections. We propose LayerDecomp, a generative framework for image layer decomposition which outputs photorealistic clean backgrounds and high-quality transparent foregrounds with faithfully preserved visual effects. To enable effective training, we first introduce a dataset preparation pipeline that automatically scales up simulated multi-layer data with synthesized visual effects. To further enhance real-world applicability, we supplement this simulated dataset with camera-captured images containing natural visual effects. Additionally, we propose a consistency loss which enforces the model to learn accurate representations for the transparent foreground layer when ground-truth annotations are not available. Our method achieves superior quality in layer decomposition, outperforming existing approaches in object removal and spatial editing tasks across several benchmarks and multiple user studies, unlocking various creative possibilities for layer-wise image editing.

Process-supervised reward models serve as a fine-grained function that provides detailed step-wise feedback to model responses, facilitating effective selection of reasoning trajectories for complex tasks.Despite its advantages, evaluation on PRMs remains less explored, especially in the multimodal domain.To address this gap, this paper first benchmarks current vision large language models (VLLMs) as two types of reward models: output reward models (ORMs) and process reward models (PRMs) on multiple vision-language benchmarks, which reveal that neither ORM nor PRM consistently outperforms across all tasks, and superior VLLMs do not necessarily yield better rewarding performance.To further advance evaluation, we introduce VILBENCH, a vision-language benchmark designed to require intensive process reward signals.Notably, Ope-nAI's GPT-4o with Chain-of-Thought (CoT) achieves only 27.3% accuracy, challenging current VLLMs.Lastly, we preliminarily showcase a promising pathway towards bridging the gap between general VLLMs and reward models-by collecting 73.6K vision-language process reward data using an enhanced treesearch algorithm, our 3B model is able to achieve an average improvement of 3.3% over standard CoT and up to 2.5% compared to its untrained counterpart on VILBENCH by selecting OpenAI o1's generations.We will release our code, model, and data at https: //ucsc-vlaa.github.io/ViLBench.

OpenAI's CLIP, released in early 2021, have long been the go-to choice of vision encoder for building multimodal foundation models. Although recent alternatives such as SigLIP have begun to challenge this status quo, to our knowledge none are fully open: their training data remains proprietary and/or their training recipes are not released. This paper fills this gap with OpenVision, a fully-open, cost-effective family of vision encoders that match or surpass the performance of OpenAI's CLIP when integrated into multimodal frameworks like LLaVA. OpenVision builds on existing works -- e.g., CLIPS for training framework and Recap-DataComp-1B for training data -- while revealing multiple key insights in enhancing encoder quality and showcasing practical benefits in advancing multimodal models. By releasing vision encoders spanning from 5.9M to 632.1M parameters, OpenVision offers practitioners a flexible trade-off between capacity and efficiency in building multimodal models: larger models deliver enhanced multimodal performance, while smaller versions enable lightweight, edge-ready multimodal deployments.

This study introduces HQ-Edit, a high-quality instruction-based image editing dataset with around 200,000 edits. Unlike prior approaches relying on attribute guidance or human feedback on building datasets, we devise a scalable data collection pipeline leveraging advanced foundation models, namely GPT-4V and DALL-E 3. To ensure its high quality, diverse examples are first collected online, expanded, and then used to create high-quality diptychs featuring input and output images with detailed text prompts, followed by precise alignment ensured through post-processing. In addition, we propose two evaluation metrics, Alignment and Coherence, to quantitatively assess the quality of image edit pairs using GPT-4V. HQ-Edits high-resolution images, rich in detail and accompanied by comprehensive editing prompts, substantially enhance the capabilities of existing image editing models. For example, an HQ-Edit finetuned InstructPix2Pix can attain state-of-the-art image editing performance, even surpassing those models fine-tuned with human-annotated data.

The vision community has started to build with the recently developed state space model, Mamba, as the new backbone for a range of tasks. This paper shows that Mamba's visual capability can be significantly enhanced through autoregressive pretraining, a direction not previously explored. Efficiency-wise, the autoregressive nature can well capitalize on the Mamba's unidirectional recurrent structure, enabling faster overall training speed compared to other training strategies like mask modeling. Performance-wise, autoregressive pretraining equips the Mamba architecture with markedly higher accuracy over its supervised-trained counterparts and, more importantly, successfully unlocks its scaling potential to large and even huge model sizes. For example, with autoregressive pretraining, a base-size Mamba attains 83.2\% ImageNet accuracy, outperforming its supervised counterpart by 2.0\%; our huge-size Mamba, the largest Vision Mamba to date, attains 85.0\% ImageNet accuracy (85.5\% when finetuned with $384\times384$ inputs), notably surpassing all other Mamba variants in vision. The code is available at \url{https://github.com/OliverRensu/ARM}.

This paper introduces MedTrinity-25M, a comprehensive, large-scale multimodal dataset for medicine, covering over 25 million images across 10 modalities with multigranular annotations for more than 65 diseases. These multigranular annotations encompass both global information, such as modality and organ detection, and local information like ROI analysis, lesion texture, and region-wise correlations. Unlike the existing multimodal datasets, which are limited by the availability of image-text pairs, we have developed the first automated pipeline that scales up multimodal data by generating multigranular visual and textual annotations in the form of image-ROI-description triplets without the need for any paired text descriptions. Specifically, data from over 30 different sources have been collected, preprocessed, and grounded using domain-specific expert models to identify ROIs related to abnormal regions. We then build a comprehensive knowledge base and prompt multimodal large language models to perform retrieval-augmented generation with the identified ROIs as guidance, resulting in multigranular textual descriptions. Compared to existing datasets, MedTrinity-25M provides the most enriched annotations, supporting a comprehensive range of multimodal tasks such as captioning and report generation, as well as vision-centric tasks like classification and segmentation. We propose LLaVA-Tri by pretraining LLaVA on MedTrinity-25M, achieving state-of-the-art performance on VQA-RAD, SLAKE, and PathVQA, surpassing representative SOTA multimodal large language models. Furthermore, MedTrinity-25M can also be utilized to support large-scale pre-training of multimodal medical AI models, contributing to the development of future foundation models in the medical domain. We will make our dataset available. The dataset is publicly available at https://yunfeixie233.github.io/MedTrinity-25M/.

Since the introduction of Vision Transformer (ViT), patchification has long been regarded as a common image pre-processing approach for plain visual architectures. By compressing the spatial size of images, this approach can effectively shorten the token sequence and reduce the computational cost of ViT-like plain architectures. In this work, we aim to thoroughly examine the information loss caused by this patchification-based compressive encoding paradigm and how it affects visual understanding. We conduct extensive patch size scaling experiments and excitedly observe an intriguing scaling law in patchification: the models can consistently benefit from decreased patch sizes and attain improved predictive performance, until it reaches the minimum patch size of 1*1, i.e., pixel tokenization. This conclusion is broadly applicable across different vision tasks, various input scales, and diverse architectures such as ViT and the recent Mamba models. Moreover, as a by-product, we discover that with smaller patches, task-specific decoder heads become less critical for dense prediction. In the experiments, we successfully scale up the visual sequence to an exceptional length of 50,176 tokens, achieving a competitive test accuracy of 84.6% with a base-sized model on the ImageNet-1k benchmark. We hope this study can provide insights and theoretical foundations for future works of building non-compressive vision models.

Web-crawled image-text pairs are inherently noisy. Prior studies demonstrate that semantically aligning and enriching textual descriptions of these pairs can significantly enhance model training across various vision-language tasks, particularly text-to-image generation. However, large-scale investigations in this area remain predominantly closed-source. Our paper aims to bridge this community effort, leveraging the powerful and $\textit{open-sourced}$ LLaMA-3, a GPT-4 level LLM. Our recaptioning pipeline is simple: first, we fine-tune a LLaMA-3-8B powered LLaVA-1.5 and then employ it to recaption ~1.3 billion images from the DataComp-1B dataset. Our empirical results confirm that this enhanced dataset, Recap-DataComp-1B, offers substantial benefits in training advanced vision-language models. For discriminative models like CLIP, we observe an average of 3.1% enhanced zero-shot performance cross four cross-modal retrieval tasks using a mixed set of the original and our captions. For generative models like text-to-image Diffusion Transformers, the generated images exhibit a significant improvement in alignment with users' text instructions, especially in following complex queries. Our project page is https://www.haqtu.me/Recap-Datacomp-1B/.

The machine learning community has witnessed a drastic change in the training pipeline, pivoted by those “foundation models” with unprecedented scales. However, the field of adversarial training is lagging behind, predominantly centered around small model sizes like ResNet-50, and tiny and low-resolution datasets like CIFAR-10. To bridge this transformation gap, this paper provides a modern re-examination with adversarial training, investigating its potential benefits when applied at scale. Additionally, we introduce an efficient and effective training strategy to enable adversarial training with giant models and web-scale data at an affordable computing cost. We denote this newly introduced framework as AdvXL. Empirical results demonstrate that AdvXL establishes new state-of-the-art robust accuracy records under AutoAttack on ImageNet-1K. For example, by training on DataComp-1B dataset, our AdvXL empowers a vanilla ViT-g model to substantially surpass the previous records of l∞-, l2-, and l1-robust accuracy by margins of 11.4%, 14.2% and 12.9%, respectively. This achievement posits AdvXL as a pioneering approach, charting a new trajectory for the efficient training of robust visual representations at significantly larger scales. Our code is available at https://github.com/UCSC-VLAA/AdvXL.

Contrastive learning has emerged as a promising paradigm for 3D open-world understanding, i.e., aligning point cloud representation to image and text embedding space individually. In this paper, we introduce Mix-Con3D, a simple yet effective method aiming to sculpt holistic 3D representation in contrastive language-image-3D pre-training. In contrast to point cloud only, we develop the 3D object-level representation from complementary perspectives, e.g., multi-view rendered images with the point cloud. Then, MixCon3D performs language-3D contrastive learning, comprehensively depicting real-world 3D objects and bolstering text alignment. Additionally, we pioneer the first thorough investigation of various training recipes for the 3D contrastive learning paradigm, building a solid baseline with improved performance. Extensive experiments conducted on three representative benchmarks reveal that our method significantly improves over the baseline, surpassing the previous state-of-the-art performance on the challenging 1,156-category Objaverse-LVIS dataset by 5.7%. The versatility of MixCon3D is showcased in applications such as text-to-3D retrieval and point cloud captioning, further evidencing its efficacy in diverse scenarios. The code is available at https://github.com/UCSC-VLAA/MixCon3D.

Deep neural networks have shown great success in representation learning. However, when learning with noisy labels (LNL), they can easily overfit and fail to generalize to new data. This paper introduces a simple and effective method, named Learning to Bootstrap (L2B), which enables models to bootstrap themselves using their own predictions without being adversely affected by erroneous pseudo-labels. It achieves this by dynamically adjusting the importance weight between real observed and generated labels, as well as between different samples through metalearning. Unlike existing instance reweighting methods, the key to our method lies in a new, versatile objective that enables implicit relabeling concurrently, leading to significant improvements without incurring additional costs. L2B offers several benefits over the baseline methods. It yields more robust models that are less susceptible to the impact of noisy labels by guiding the bootstrapping procedure more effectively. It better exploits the valuable information contained in corrupted instances by adapting the weights of both instances and labels. Furthermore, L2B is compatible with existing LNL methods and delivers competitive results spanning natural and medical imaging tasks including classification and segmentation under both synthetic and real-world noise. Extensive experiments demonstrate that our method effectively mitigates the challenges of noisy labels, often necessitating few to no validation samples, and is well generalized to other tasks such as image segmentation. This not only positions it as a robust complement to existing LNL techniques but also underscores its practical applicability. The code and models are available at https://github.com/yuyinzhou/12b.

Despite the impressive advancements achieved through vision-and-language pretraining, it remains unclear whether this joint learning paradigm can help understand each individual modality.In this work, we conduct a comparative analysis of the visual representations in existing vision-and-language models and visiononly models by probing a broad range of tasks, aiming to assess the quality of the learned representations in a nuanced manner.Interestingly, our empirical observations suggest that visionand-language models are better at label prediction tasks like object and attribute prediction, while vision-only models are stronger at dense prediction tasks that require more localized information.We hope our study sheds light on the role of language in visual learning, and serves as an empirical guide for various pretrained models.Code will be released at https: //github.com/Lizw14/visual_probing.

This paper introduces an efficient strategy to transform Large Language Models (LLMs) into Multi-Modal Large Language Models. By conceptualizing this transformation as a domain adaptation process, \ie, transitioning from text understanding to embracing multiple modalities, we intriguingly note that, within each attention block, tuning LayerNorm suffices to yield strong performance. Moreover, when benchmarked against other tuning approaches like full parameter finetuning or LoRA, its benefits on efficiency are substantial. For example, when compared to LoRA on a 13B model scale, performance can be enhanced by an average of over 20\% across five multi-modal tasks, and meanwhile, results in a significant reduction of trainable parameters by 41.9\% and a decrease in GPU memory usage by 17.6\%. On top of this LayerNorm strategy, we showcase that selectively tuning only with conversational data can improve efficiency further. Beyond these empirical outcomes, we provide a comprehensive analysis to explore the role of LayerNorm in adapting LLMs to the multi-modal domain and improving the expressive power of the model.

This paper enhances image-GPT (iGPT), one of the pioneering works that introduce autoregressive pretraining to predict the next pixels for visual representation learning. Two simple yet essential changes are made. First, we shift the prediction target from raw pixels to semantic tokens, enabling a higher-level understanding of visual content. Second, we supplement the autoregressive modeling by instructing the model to predict not only the next tokens but also the visible tokens. This pipeline is particularly effective when semantic tokens are encoded by discriminatively trained models, such as CLIP. We introduce this novel approach as D-iGPT. Extensive experiments showcase that D-iGPT excels as a strong learner of visual representations: A notable achievement is its compelling performance on the ImageNet-1K dataset — by training on publicly available datasets, D-iGPT unprecedentedly achieves 90.0% top-1 accuracy with a vanilla ViT-H. Additionally, D-iGPT shows strong generalization on the downstream task. Code is available at https://github.com/OliverRensu/D-iGPT.

Yunchao Zhang, Zonglin Di, Kaiwen Zhou, Cihang Xie, Xin Wang. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Current benchmarks for assessing vision-language models (VLMs) often focus on their perception or problem-solving capabilities and neglect other critical aspects such as fairness, multilinguality, or toxicity. Furthermore, they differ in their evaluation procedures and the scope of the evaluation, making it difficult to compare models. To address these issues, we extend the HELM framework to VLMs to present the Holistic Evaluation of Vision Language Models (VHELM). VHELM aggregates various datasets to cover one or more of the 9 aspects: visual perception, knowledge, reasoning, bias, fairness, multilinguality, robustness, toxicity, and safety. In doing so, we produce a comprehensive, multi-dimensional view of the capabilities of the VLMs across these important factors. In addition, we standardize the standard inference parameters, methods of prompting, and evaluation metrics to enable fair comparisons across models. Our framework is designed to be lightweight and automatic so that evaluation runs are cheap and fast. Our initial run evaluates 22 VLMs on 21 existing datasets to provide a holistic snapshot of the models. We uncover new key findings, such as the fact that efficiency-focused models (e.g., Claude 3 Haiku or Gemini 1.5 Flash) perform significantly worse than their full models (e.g., Claude 3 Opus or Gemini 1.5 Pro) on the bias benchmark but not when evaluated on the other aspects. For transparency, we release the raw model generations and complete results on our website at https://crfm.stanford.edu/helm/vhelm/v2.0.1. VHELM is intended to be a living benchmark, and we hope to continue adding new datasets and models over time.

CRATE, a white-box transformer architecture designed to learn compressed and sparse representations, offers an intriguing alternative to standard vision transformers (ViTs) due to its inherent mathematical interpretability. Despite extensive investigations into the scaling behaviors of language and vision transformers, the scalability of CRATE remains an open question which this paper aims to address. Specifically, we propose CRATE-$\alpha$, featuring strategic yet minimal modifications to the sparse coding block in the CRATE architecture design, and a light training recipe designed to improve the scalability of CRATE. Through extensive experiments, we demonstrate that CRATE-$\alpha$ can effectively scale with larger model sizes and datasets. For example, our CRATE-$\alpha$-B substantially outperforms the prior best CRATE-B model accuracy on ImageNet classification by 3.7%, achieving an accuracy of 83.2%. Meanwhile, when scaling further, our CRATE-$\alpha$-L obtains an ImageNet classification accuracy of 85.1%. More notably, these model performance improvements are achieved while preserving, and potentially even enhancing the interpretability of learned CRATE models, as we demonstrate through showing that the learned token representations of increasingly larger trained CRATE-$\alpha$ models yield increasingly higher-quality unsupervised object segmentation of images.

By harnessing the latest advances in deep learning, image-to-image translation architectures have recently achieved impressive capabilities. Unfortunately, the growing representational power of these architectures has prominent unethical uses. Among these, the threats of (1) face manipulation ("DeepFakes") used for misinformation or pornographic use (2) "DeepNude" manipulations of body images to remove clothes from individuals, etc. Several works tackle the task of disrupting such image translation networks by inserting imperceptible adversarial attacks into the input image. Nevertheless, these works have limitations that may result in disruptions that are not practical in the real world. Specifically, most works generate disruptions in a white-box scenario, assuming perfect knowledge about the image translation network. The few remaining works that assume a black-box scenario require a large number of queries to successfully disrupt the adversary's image translation network. In this work we propose Leaking Transferable Perturbations (LTP), an algorithm that significantly reduces the number of queries needed to disrupt an image translation network by dynamically re-purposing previous disruptions into new query efficient disruptions.

This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps. Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves 84.0% ImageNet top-1 accuracy, outperforming the baseline of directly distilling afinetuned ViT-L by 1.2%. More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with 95% masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of 83.6%; surprisingly, it can still secure 82.4% top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code and models are publicly available at https://github.com/UCSC-VLAA/DMAE.

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

3D perception based on the representations learned from multi-camera bird’s-eye-view (BEV) is trending as cameras are cost-effective for mass production in autonomous driving industry. However, there exists a distinct performance gap between multi-camera BEV and LiDAR based 3D object detection. One key reason is that LiDAR captures accurate depth and other geometry measurements, while it is notoriously challenging to infer such 3D information from merely image input. In this work, we propose to boost the representation learning of a multi-camera BEV based student detector by training it to imitate the features of a well-trained LiDAR based teacher detector. We propose effective balancing strategy to enforce the student to focus on learning the crucial features from the teacher, and generalize knowledge transfer to multi-scale layers with temporal fusion. We conduct extensive evaluations on multiple representative models of multi-camera BEV. Experiments reveal that our approach renders significant improvement over the student models, leading to the state-of-the-art performance on the popular benchmark nuScenes.

Video-language pre-training is crucial for learning powerful multi-modal representation. However, it typically requires a massive amount of computation. In this paper, we develop SMAUG, an efficient pre-training framework for video-language models. The foundation component in SMAUG is masked autoencoders. Different from prior works which only mask textual inputs, our masking strategy considers both visual and textual modalities, providing a better cross-modal alignment and saving more pre-training costs. On top of that, we introduce a space-time token sparsification module, which leverages context information to further select only "important" spatial regions and temporal frames for pre-training. Coupling all these designs allows our method to enjoy both competitive performances on text-to-video retrieval and video question answering tasks, and much less pre-training costs by 1.9× or more. For example, our SMAUG only needs ~50 NVIDIA A6000 GPU hours for pre-training to attain competitive performances on these two video-language tasks across six popular benchmarks.

CLIP, one of the pioneering foundation models that connect images and text, has enabled many recent breakthroughs in computer vision. However, its associated training cost is prohibitively high, imposing a significant barrier to its widespread exploration. In this paper, we present a surprising finding that there exists an inverse scaling law for CLIP training, whereby the larger the image/text encoders used, the shorter the sequence length of image/text tokens that can be applied in training. Moreover, we showcase that the strategy for reducing image/text token length plays a crucial role in determining the quality of this scaling law. As a result of this finding, we are able to successfully train CLIP even with limited computational resources. For example, using 8 A100 GPUs, our CLIP models achieve zero-shot top-1 ImageNet-1k accuracies of 63.2% in ~2 days, 67.8% in ~3 days, and 69.3% in ~4 days. Our method also works well when scaling up --- with G/14, we register a new record of 83.0% ImageNet-1k zero-shot accuracy, and meanwhile accelerate the training by ~33x compared to its OpenCLIP counterpart. By reducing the computation barrier associated with CLIP, we hope to inspire more research in this field, particularly from academics. Our code is available at https://github.com/UCSC-VLAA/CLIPA.

Batch normalization (BN) is a fundamental unit in modern deep neural networks. However, BN and its variants focus on normalization statistics but neglect the recovery step that uses linear transformation to improve the capacity of fitting complex data distributions. In this paper, we demonstrate that the recovery step can be improved by aggregating the neighborhood of each neuron rather than just considering a single neuron. Specifically, we propose a simple yet effective method named batch normalization with enhanced linear transformation (BNET) to embed spatial contextual information and improve representation ability. BNET can be easily implemented using the depth-wise convolution and seamlessly transplanted into existing architectures with BN. To our best knowledge, BNET is the first attempt to enhance the recovery step for BN. Furthermore, BN is interpreted as a special case of BNET from both spatial and spectral views. Experimental results demonstrate that BNET achieves consistent performance gains based on various backbones in a wide range of visual tasks. Moreover, BNET can accelerate the convergence of network training and enhance spatial information by assigning important neurons with large weights accordingly.

Data mixing (e.g., Mixup, Cutmix, ResizeMix) is an essential component for advancing recognition models. In this paper, we focus on studying its effectiveness in the self-supervised setting. By noticing the mixed images that share the same source images are intrinsically related to each other, we hereby propose SDMP, short for Simple Data Mixing Prior, to capture this straightforward yet essential prior, and position such mixed images as additional positive pairs to facilitate self-supervised representation learning. Our experiments verify that the proposed SDMP enables data mixing to help a set of self-supervised learning frameworks (e.g., MoCo) achieve better accuracy and out-of-distribution robustness. More notably, our SDMP is the first method that successfully leverages data mixing to improve (rather than hurt) the performance of Vision Transformers in the self-supervised setting. Code is publicly available at https://github.com/OliverRensu/SDMP.

Most machine learning models are validated and tested on fixed datasets. This can give an incomplete picture of the capabilities and weaknesses of the model. Such weaknesses can be revealed at test time in the real world. The risks involved in such failures can be loss of profits, loss of time or even loss of life in certain critical applications. In order to alleviate this issue, simulators can be controlled in a finegrained manner using interpretable parameters to explore the semantic image manifold. In this work, we propose a framework for learning how to test machine learning algorithms using simulators in an adversarial manner in order to find weaknesses in the model before deploying it in critical scenarios. We apply this method in a face recognition setup. We show that certain weaknesses of models trained on real data can be discovered using simulated samples. Using our proposed method, we can find adversarial synthetic faces that fool contemporary face recognition models. This demonstrates the fact that these models have weaknesses that are not measured by commonly used validation datasets. We hypothesize that this type of adversarial examples are not isolated, but usually lie in connected spaces in the latent space of the simulator. We present a method to find these adversarial regions as opposed to the typical adversarial points found in the adversarial example literature.

The score-based query attacks (SQAs) pose practical threats to deep neural networks by crafting adversarial perturbations within dozens of queries, only using the model's output scores. Nonetheless, we note that if the loss trend of the outputs is slightly perturbed, SQAs could be easily misled and thereby become much less effective. Following this idea, we propose a novel defense, namely Adversarial Attack on Attackers (AAA), to confound SQAs towards incorrect attack directions by slightly modifying the output logits. In this way, (1) SQAs are prevented regardless of the model's worst-case robustness; (2) the original model predictions are hardly changed, i.e., no degradation on clean accuracy; (3) the calibration of confidence scores can be improved simultaneously. Extensive experiments are provided to verify the above advantages. For example, by setting $\ell_\infty=8/255$ on CIFAR-10, our proposed AAA helps WideResNet-28 secure 80.59% accuracy under Square attack (2500 queries), while the best prior defense (i.e., adversarial training) only attains 67.44%. Since AAA attacks SQA's general greedy strategy, such advantages of AAA over 8 defenses can be consistently observed on 8 CIFAR-10/ImageNet models under 6 SQAs, using different attack targets, bounds, norms, losses, and strategies. Moreover, AAA calibrates better without hurting the accuracy. Our code is available at https://github.com/Sizhe-Chen/AAA.

Modern deep neural networks tend to be evaluated on static test sets. One shortcoming of this is the fact that these deep neural networks cannot be easily evaluated for robustness issues with respect to specific scene variations. For example, it is hard to study the robustness of these networks to variations of object scale, object pose, scene lighting and 3D occlusions. The main reason is that collecting real datasets with fine-grained naturalistic variations of sufficient scale can be extremely time-consuming and expensive. In this work, we present Counterfactual Simulation Testing, a counterfactual framework that allows us to study the robustness of neural networks with respect to some of these naturalistic variations by building realistic synthetic scenes that allow us to ask counterfactual questions to the models, ultimately providing answers to questions such as "Would your classification still be correct if the object were viewed from the top?" or "Would your classification still be correct if the object were partially occluded by another object?". Our method allows for a fair comparison of the robustness of recently released, state-of-the-art Convolutional Neural Networks and Vision Transformers, with respect to these naturalistic variations. We find evidence that ConvNext is more robust to pose and scale variations than Swin, that ConvNext generalizes better to our simulated domain and that Swin handles partial occlusion better than ConvNext. We also find that robustness for all networks improves with network scale and with data scale and variety. We release the Naturalistic Variation Object Dataset (NVD), a large simulated dataset of 272k images of everyday objects with naturalistic variations such as object pose, scale, viewpoint, lighting and occlusions. Project page: https://counterfactualsimulation.github.io

Data augmentation has become a de facto component for training high-performance deep image classifiers, but its potential is under-explored for object detection. Noting that most state-of-the-art object detectors benefit from fine-tuning a pre-trained classifier, we first study how the classifiers’ gains from various data augmentations transfer to object detection. The results are discouraging; the gains diminish after fine-tuning in terms of either accuracy or robustness. This work instead augments the fine-tuning stage for object detectors by exploring adversarial examples, which can be viewed as a model-dependent data augmentation. Our method dynamically selects the stronger adversarial images sourced from a detector’s classification and localization branches and evolves with the detector to ensure the augmentation policy stays current and relevant. This model-dependent augmentation generalizes to different object detectors better than AutoAugment, a model-agnostic augmentation policy searched based on one particular detector. Our approach boosts the performance of state-of-the-art EfficientDets by +1.1 mAP on the COCO object detection benchmark. It also improves the detectors’ robustness against natural distortions by +3.8 mAP and against domain shift by +1.3 mAP.

While neural symbolic methods demonstrate impressive performance in visual question answering on synthetic images, their performance suffers on real images. We identify that the long-tail distribution of visual concepts and unequal importance of reasoning steps in real data are the two key obstacles that limit the models’ real-world potentials. To address these challenges, we propose a new paradigm, Calibrating Concepts and Operations (CCO), which enables neural symbolic models to capture underlying data characteristics and to reason with hierarchical importance. Specifically, we introduce an executor with learnable concept embedding magnitudes for handling distribution imbalance, and an operation calibrator for highlighting important operations and suppressing redundant ones.Our experiments show CCO substantially boosts the performance of neural symbolic methods on real images. By evaluating models on the real world dataset GQA, CCO helps the neural symbolic method NSCL outperforms its vanilla counterpart by 9.1% (from 47.0% to 56.1%); this result also largely reduces the performance gap between symbolic and non-symbolic methods. Additionally, we create a perturbed test set for better understanding and analyzing model performance on real images. Code is available at https://lizw14.github.io/project/ccosr.

Transformer emerges as a powerful tool for visual recognition. In addition to demonstrating competitive performance on a broad range of visual benchmarks, recent works also argue that Transformers are much more robust than Convolutions Neural Networks (CNNs). Nonetheless, surprisingly, we find these conclusions are drawn from unfair experimental settings, where Transformers and CNNs are compared at different scales and are applied with distinct training frameworks. In this paper, we aim to provide the first fair & in-depth comparisons between Transformers and CNNs, focusing on robustness evaluations. With our unified training setup, we first challenge the previous belief that Transformers outshine CNNs when measuring adversarial robustness. More surprisingly, we find CNNs can easily be as robust as Transformers on defending against adversarial attacks, if they properly adopt Transformers' training recipes. While regarding generalization on out-of-distribution samples, we show pre-training on (external) large-scale datasets is not a fundamental request for enabling Transformers to achieve better performance than CNNs. Moreover, our ablations suggest such stronger generalization is largely benefited by the Transformer's self-attention-like architectures per se, rather than by other training setups. We hope this work can help the community better understand and benchmark the robustness of Transformers and CNNs. The code and models are publicly available at: https://github.com/ytongbai/ViTs-vs-CNNs.

Recent development of adversarial attacks has proven that ensemble-based methods outperform traditional, non-ensemble ones in black-box attack. However, as it is computationally prohibitive to acquire a family of diverse models, these methods achieve inferior performance constrained by the limited number of models to be ensembled.In this paper, we propose Ghost Networks to improve the transferability of adversarial examples. The critical principle of ghost networks is to apply feature-level perturbations to an existing model to potentially create a huge set of diverse models. After that, models are subsequently fused by longitudinal ensemble. Extensive experimental results suggest that the number of networks is essential for improving the transferability of adversarial examples, but it is less necessary to independently train different networks and ensemble them in an intensive aggregation way. Instead, our work can be used as a computationally cheap and easily applied plug-in to improve adversarial approaches both in single-model and multi-model attack, compatible with residual and non-residual networks. By reproducing the NeurIPS 2017 adversarial competition, our method outperforms the No.1 attack submission by a large margin, demonstrating its effectiveness and efficiency. Code is available at https://github.com/LiYingwei/ghost-network.

In this paper, we study physical adversarial attacks on object detectors in the wild. Previous works mostly craft instance-dependent perturbations only for rigid or planar objects. To this end, we propose to learn an adversarial pattern to effectively attack all instances belonging to the same object category, referred to as Universal Physical Camouflage Attack (UPC). Concretely, UPC crafts camouflage by jointly fooling the region proposal network, as well as misleading the classifier and the regressor to output errors. In order to make UPC effective for non-rigid or non-planar objects, we introduce a set of transformations for mimicking deformable properties. We additionally impose optimization constraint to make generated patterns look natural to human observers. To fairly evaluate the effectiveness of different physical-world attacks, we present the first standardized virtual database, AttackScenes, which simulates the real 3D world in a controllable and reproducible environment. Extensive experiments suggest the superiority of our proposed UPC compared with existing physical adversarial attackers not only in virtual environments (AttackScenes), but also in real-world physical environments.

Non-Local (NL) blocks have been widely studied in various vision tasks. However, it has been rarely explored to embed the NL blocks in mobile neural networks, mainly due to the following challenges: 1) NL blocks generally have heavy computation cost which makes it difficult to be applied in applications where computational resources are limited, and 2) it is an open problem to discover an optimal configuration to embed NL blocks into mobile neural networks. We propose AutoNL to overcome the above two obstacles. Firstly, we propose a Lightweight Non-Local (LightNL) block by squeezing the transformation operations and incorporating compact features. With the novel design choices, the proposed LightNL block is 400 times computationally cheaper} than its conventional counterpart without sacrificing the performance. Secondly, by relaxing the structure of the LightNL block to be differentiable during training, we propose an efficient neural architecture search algorithm to learn an optimal configuration of LightNL blocks in an end-to-end manner. Notably, using only 32 GPU hours, the searched AutoNL model achieves 77.7% top-1 accuracy on ImageNet under a typical mobile setting (350M FLOPs), significantly outperforming previous mobile models including MobileNetV2 (+5.7%), FBNet (+2.8%) and MnasNet (+2.1%). Code and models are available at https://github.com/LiYingwei/AutoNL.

Adversarial examples are commonly viewed as a threat to ConvNets. Here we present an opposite perspective: adversarial examples can be used to improve image recognition models if harnessed in the right manner. We propose AdvProp, an enhanced adversarial training scheme which treats adversarial examples as additional examples, to prevent overfitting. Key to our method is the usage of a separate auxiliary batch norm for adversarial examples, as they have different underlying distributions to normal examples. We show that AdvProp improves a wide range of models on various image recognition tasks and performs better when the models are bigger. For instance, by applying AdvProp to the latest EfficientNet-B7 [28] on ImageNet, we achieve significant improvements on ImageNet (+0.7%), ImageNet-C (+6.5%), ImageNet-A (+7.0%), Stylized-ImageNet (+4.8%). With an enhanced EfficientNet-B8, our method achieves the state-of-the-art 85.5% ImageNet top-1 accuracy without extra data. This result even surpasses the best model in [20] which is trained with 3.5B Instagram images (~3000X more than ImageNet) and ~9.4X more parameters. Code and models will be made publicly available.

Adversarial attacks to image classification systems present challenges to convolutional networks and opportunities for understanding them. This study suggests that adversarial perturbations on images lead to noise in the features constructed by these networks. Motivated by this observation, we develop new network architectures that increase adversarial robustness by performing feature denoising. Specifically, our networks contain blocks that denoise the features using non-local means or other filters; the entire networks are trained end-to-end. When combined with adversarial training, our feature denoising networks substantially improve the state-of-the-art in adversarial robustness in both white-box and black-box attack settings. On ImageNet, under 10-iteration PGD white-box attacks where prior art has 27.9% accuracy, our method achieves 55.7%; even under extreme 2000-iteration PGD white-box attacks, our method secures 42.6% accuracy. Our method was ranked first in Competition on Adversarial Attacks and Defenses (CAAD) 2018 --- it achieved 50.6% classification accuracy on a secret, ImageNet-like test dataset against 48 unknown attackers, surpassing the runner-up approach by ~10%. Code is available at https://github.com/facebookresearch/ImageNet-Adversarial-Training.

Though CNNs have achieved the state-of-the-art performance on various vision tasks, they are vulnerable to adversarial examples --- crafted by adding human-imperceptible perturbations to clean images. However, most of the existing adversarial attacks only achieve relatively low success rates under the challenging black-box setting, where the attackers have no knowledge of the model structure and parameters. To this end, we propose to improve the transferability of adversarial examples by creating diverse input patterns. Instead of only using the original images to generate adversarial examples, our method applies random transformations to the input images at each iteration. Extensive experiments on ImageNet show that the proposed attack method can generate adversarial examples that transfer much better to different networks than existing baselines. By evaluating our method against top defense solutions and official baselines from NIPS 2017 adversarial competition, the enhanced attack reaches an average success rate of 73.0%, which outperforms the top-1 attack submission in the NIPS competition by a large margin of 6.6%. We hope that our proposed attack strategy can serve as a strong benchmark baseline for evaluating the robustness of networks to adversaries and the effectiveness of different defense methods in the future. Code is available at https://github.com/cihangxie/DI-2-FGSM.

We propose a novel single shot object detection network named Detection with Enriched Semantics (DES). Our motivation is to enrich the semantics of object detection features within a typical deep detector, by a semantic segmentation branch and a global activation module. The segmentation branch is supervised by weak segmentation ground-truth, i.e., no extra annotation is required. In conjunction with that, we employ a global activation module which learns relationship between channels and object classes in a self-supervised manner. Comprehensive experimental results on both PASCAL VOC and MS COCO detection datasets demonstrate the effectiveness of the proposed method. In particular, with a VGG16 based DES, we achieve an mAP of 81.7 on VOC2007 test and an mAP of 32.8 on COCO test-dev with an inference speed of 31.5 milliseconds per image on a Titan Xp GPU. With a lower resolution version, we achieve an mAP of 79.7 on VOC2007 with an inference speed of 13.0 milliseconds per image.

In this paper, we study the task of detecting semantic parts of an object, e.g., a wheel of a car, under partial occlusion. We propose that all models should be trained without seeing occlusions while being able to transfer the learned knowledge to deal with occlusions. This setting alleviates the difficulty in collecting an exponentially large dataset to cover occlusion patterns and is more essential. In this scenario, the proposal-based deep networks, like RCNN-series, often produce unsatisfactory results, because both the proposal extraction and classification stages may be confused by the irrelevant occluders. To address this, [25] proposed a voting mechanism that combines multiple local visual cues to detect semantic parts. The semantic parts can still be detected even though some visual cues are missing due to occlusions. However, this method is manually-designed, thus is hard to be optimized in an end-to-end manner. In this paper, we present DeepVoting, which incorporates the robustness shown by [25] into a deep network, so that the whole pipeline can be jointly optimized. Specifically, it adds two layers after the intermediate features of a deep network, e.g., the pool-4 layer of VGGNet. The first layer extracts the evidence of local visual cues, and the second layer performs a voting mechanism by utilizing the spatial relationship between visual cues and semantic parts. We also propose an improved version DeepVoting+ by learning visual cues from context outside objects. In experiments, DeepVoting achieves significantly better performance than several baseline methods, including Faster-RCNN, for semantic part detection under occlusion. In addition, DeepVoting enjoys explainability as the detection results can be diagnosed via looking up the voting cues.

It has been well demonstrated that adversarial examples, i.e., natural images with visually imperceptible perturbations added, cause deep networks to fail on image classification. In this paper, we extend adversarial examples to semantic segmentation and object detection which are much more difficult. Our observation is that both segmentation and detection are based on classifying multiple targets on an image (e.g., the target is a pixel or a receptive field in segmentation, and an object proposal in detection). This inspires us to optimize a loss function over a set of targets for generating adversarial perturbations. Based on this, we propose a novel algorithm named Dense Adversary Generation (DAG), which applies to the state-of-the-art networks for segmentation and detection. We find that the adversarial perturbations can be transferred across networks with different training data, based on different architectures, and even for different recognition tasks. In particular, the transfer ability across networks with the same architecture is more significant than in other cases. Besides, we show that summing up heterogeneous perturbations often leads to better transfer performance, which provides an effective method of black-box adversarial attack.

Abstract We consider a cognitive radio network which coexists with multiple primary users (PUs) and secondary users (SUs) transmit over time‐varying channels. In this scenario, one problem of the existing work is the poor performances of throughput and fairness due to variances of SUs' channel conditions and PUs' traffic patterns. To solve this problem, we propose a novel prediction‐based MAC‐layer sensing algorithm. In the proposed algorithm, the SUs' channel quality information and the probability of the licensed channel being idle are predicted. Through the earlier predicted information, we schedule the SUs to sense and transmit on different licensed channels. Specifically, multiple significant factors, including network throughput and fairness, are jointly considered in the proposed algorithm. Then, we formulate the prediction‐based sensing scheduling problem as an optimization problem and solve it with the Hungarian algorithm in polynomial time. Simulation results show that the proposed prediction‐based sensing scheduling algorithm could achieve a good tradeoff between network throughput and fairness among SUs. Copyright © 2014 John Wiley & Sons, Ltd.