Youngwan Lee

Youngwan Lee, Jeffrey Willette, Jonghee Kim et al.

Masked image modeling (MIM) has become a popular strategy for self-supervised learning~(SSL) of visual representations with Vision Transformers. A representative MIM model, the masked auto-encoder (MAE), randomly masks a subset of image patches and reconstructs the masked patches given the unmasked patches. Concurrently, many recent works in self-supervised learning utilize the student/teacher paradigm which provides the student with an additional target based on the output of a teacher composed of an exponential moving average (EMA) of previous students. Although common, relatively little is known about the dynamics of the interaction between the student and teacher. Through analysis on a simple linear model, we find that the teacher conditionally removes previous gradient directions based on feature similarities which effectively acts as a conditional momentum regularizer. From this analysis, we present a simple SSL method, the Reconstruction-Consistent Masked Auto-Encoder (RC-MAE) by adding an EMA teacher to MAE. We find that RC-MAE converges faster and requires less memory usage than state-of-the-art self-distillation methods during pre-training, which may provide a way to enhance the practicality of prohibitively expensive self-supervised learning of Vision Transformer models. Additionally, we show that RC-MAE achieves more robustness and better performance compared to MAE on downstream tasks such as ImageNet-1K classification, object detection, and instance segmentation.

Sunil Hwang, Jaehong Yoon, Youngwan Lee et al.

Masked Video Autoencoder (MVA) approaches have demonstrated their potential by significantly outperforming previous video representation learning methods. However, they waste an excessive amount of computations and memory in predicting uninformative tokens/frames due to random masking strategies. (e.g., over 16 nodes with 128 NVIDIA A100 GPUs). To resolve this issue, we exploit the unequal information density among the patches in videos and propose EVEREST, a surprisingly efficient MVA approach for video representation learning that finds tokens containing rich motion features and discards uninformative ones during both pre-training and fine-tuning. We further present an information-intensive frame selection strategy that allows the model to focus on informative and causal frames with minimal redundancy. Our method significantly reduces the computation and memory requirements of MVA, enabling the pre-training and fine-tuning on a single machine with 8 GPUs while achieving comparable performance to computation- and memory-heavy baselines on multiple benchmarks and the uncurated Ego4D dataset. We hope that our work contributes to reducing the barrier to further research on video understanding.

Youngwan Lee, Kwanyong Park, Yoorhim Cho et al.

As text-to-image (T2I) synthesis models increase in size, they demand higher inference costs due to the need for more expensive GPUs with larger memory, which makes it challenging to reproduce these models in addition to the restricted access to training datasets. Our study aims to reduce these inference costs and explores how far the generative capabilities of T2I models can be extended using only publicly available datasets and open-source models. To this end, by using the de facto standard text-to-image model, Stable Diffusion XL (SDXL), we present three key practices in building an efficient T2I model: (1) Knowledge distillation: we explore how to effectively distill the generation capability of SDXL into an efficient U-Net and find that self-attention is the most crucial part. (2) Data: despite fewer samples, high-resolution images with rich captions are more crucial than a larger number of low-resolution images with short captions. (3) Teacher: Step-distilled Teacher allows T2I models to reduce the noising steps. Based on these findings, we build two types of efficient text-to-image models, called KOALA-Turbo &-Lightning, with two compact U-Nets (1B & 700M), reducing the model size up to 54% and 69% of the SDXL U-Net. In particular, the KOALA-Lightning-700M is 4x faster than SDXL while still maintaining satisfactory generation quality. Moreover, unlike SDXL, our KOALA models can generate 1024px high-resolution images on consumer-grade GPUs with 8GB of VRAMs (3060Ti). We believe that our KOALA models will have a significant practical impact, serving as cost-effective alternatives to SDXL for academic researchers and general users in resource-constrained environments.

Kangsan Kim, Geon Park, Youngwan Lee et al.

Recent advancements in video large multimodal models (LMMs) have significantly improved their video understanding and reasoning capabilities. However, their performance drops on out-of-distribution (OOD) tasks that are underrepresented in training data. Traditional methods like fine-tuning on OOD datasets are impractical due to high computational costs. While In-context learning (ICL) with demonstration examples has shown promising generalization performance in language tasks and image-language tasks without fine-tuning, applying ICL to video-language tasks faces challenges due to the limited context length in Video LMMs, as videos require longer token lengths. To address these issues, we propose VideoICL, a novel video in-context learning framework for OOD tasks that introduces a similarity-based relevant example selection strategy and a confidence-based iterative inference approach. This allows to select the most relevant examples and rank them based on similarity, to be used for inference. If the generated response has low confidence, our framework selects new examples and performs inference again, iteratively refining the results until a high-confidence response is obtained. This approach improves OOD video understanding performance by extending effective context length without incurring high costs. The experimental results on multiple benchmarks demonstrate significant performance gains, especially in domain-specific scenarios, laying the groundwork for broader video comprehension applications. Code will be released at https://github.com/KangsanKim07/VideoICL

Kangsan Kim, Yanlai Yang, Suji Kim et al.

As embodied models become powerful, humans will collaborate with multiple embodied AI agents at their workplace or home in the future. To ensure better communication between human users and the multi-agent system, it is crucial to interpret incoming information from agents in parallel and refer to the appropriate context for each query. Existing challenges include effectively compressing and communicating high volumes of individual sensory inputs in the form of video and correctly aggregating multiple egocentric videos to construct system-level memory. In this work, we first formally define a novel problem of understanding multiple long-horizon egocentric videos simultaneously collected from embodied agents. To facilitate research in this direction, we introduce MultiAgent-EgoQA (MA-EgoQA), a benchmark designed to systemically evaluate existing models in our scenario. MA-EgoQA provides 1.7k questions unique to multiple egocentric streams, spanning five categories: social interaction, task coordination, theory-of-mind, temporal reasoning, and environmental interaction. We further propose a simple baseline model for MA-EgoQA named EgoMAS, which leverages shared memory across embodied agents and agent-wise dynamic retrieval. Through comprehensive evaluation across diverse baselines and EgoMAS on MA-EgoQA, we find that current approaches are unable to effectively handle multiple egocentric streams, highlighting the need for future advances in system-level understanding across the agents. The code and benchmark are available at https://ma-egoqa.github.io.

Youngwan Lee, Soojin Jang, Yoorhim Cho et al.

Spatial reasoning is foundational for Vision-Language Models (VLMs), particularly when deployed as Vision-Language-Action (VLA) agents in physical environments. However, existing benchmarks predominantly focus on elementary, single-hop relations, neglecting the multi-hop compositional reasoning and precise visual grounding essential for real-world scenarios. To address this, we introduce MultihopSpatial, offering three key contributions: (1) A comprehensive benchmark designed for multi-hop and compositional spatial reasoning, featuring 1- to 3-hop complex queries across diverse spatial perspectives. (2) Acc@50IoU, a complementary metric that simultaneously evaluates reasoning and visual grounding by requiring both answer selection and precise bounding box prediction - capabilities vital for robust VLA deployment. (3) MultihopSpatial-Train, a dedicated large-scale training corpus to foster spatial intelligence. Extensive evaluation of 37 state-of-the-art VLMs yields eight key insights, revealing that compositional spatial reasoning remains a formidable challenge. Finally, we demonstrate that reinforcement learning post-training on our corpus enhances both intrinsic VLM spatial reasoning and downstream embodied manipulation performance.

Youngwan Lee, Jongyoul Park

We propose a simple yet efficient anchor-free instance segmentation, called CenterMask, that adds a novel spatial attention-guided mask (SAG-Mask) branch to anchor-free one stage object detector (FCOS) in the same vein with Mask R-CNN. Plugged into the FCOS object detector, the SAG-Mask branch predicts a segmentation mask on each box with the spatial attention map that helps to focus on informative pixels and suppress noise. We also present an improved backbone networks, VoVNetV2, with two effective strategies: (1) residual connection for alleviating the optimization problem of larger VoVNet \cite{lee2019energy} and (2) effective Squeeze-Excitation (eSE) dealing with the channel information loss problem of original SE. With SAG-Mask and VoVNetV2, we deign CenterMask and CenterMask-Lite that are targeted to large and small models, respectively. Using the same ResNet-101-FPN backbone, CenterMask achieves 38.3%, surpassing all previous state-of-the-art methods while at a much faster speed. CenterMask-Lite also outperforms the state-of-the-art by large margins at over 35fps on Titan Xp. We hope that CenterMask and VoVNetV2 can serve as a solid baseline of real-time instance segmentation and backbone network for various vision tasks, respectively. The Code is available at https://github.com/youngwanLEE/CenterMask.

Sungjun Ahn, Hyun-Jeong Yim, Youngwan Lee et al.

This paper introduces a media service model that exploits artificial intelligence (AI) video generators at the receive end. This proposal deviates from the traditional multimedia ecosystem, completely relying on in-house production, by shifting part of the content creation onto the receiver. We bring a semantic process into the framework, allowing the distribution network to provide service elements that prompt the content generator, rather than distributing encoded data of fully finished programs. The service elements include fine-tailored text descriptions, lightweight image data of some objects, or application programming interfaces, comprehensively referred to as semantic sources, and the user terminal translates the received semantic data into video frames. Empowered by the random nature of generative AI, the users could then experience super-personalized services accordingly. The proposed idea incorporates the situations in which the user receives different service providers' element packages; a sequence of packages over time, or multiple packages at the same time. Given promised in-context coherence and content integrity, the combinatory dynamics will amplify the service diversity, allowing the users to always chance upon new experiences. This work particularly aims at short-form videos and advertisements, which the users would easily feel fatigued by seeing the same frame sequence every time. In those use cases, the content provider's role will be recast as scripting semantic sources, transformed from a thorough producer. Overall, this work explores a new form of media ecosystem facilitated by receiver-embedded generative models, featuring both random content dynamics and enhanced delivery efficiency simultaneously.

Youngwan Lee, Kangsan Kim, Kwanyong Park et al.

Despite emerging efforts to enhance the safety of Vision-Language Models (VLMs), current approaches face two main shortcomings. 1) Existing safety-tuning datasets and benchmarks only partially consider how image-text interactions can yield harmful content, often overlooking contextually unsafe outcomes from seemingly benign pairs. This narrow coverage leaves VLMs vulnerable to jailbreak attacks in unseen configurations. 2) Prior methods rely primarily on data-centric tuning, with limited architectural innovations to intrinsically strengthen safety. We address these gaps by introducing a holistic safety dataset and benchmark, \textbf{HoliSafe}, that spans all five safe/unsafe image-text combinations, providing a more robust basis for both training and evaluation (HoliSafe-Bench). We further propose a novel modular framework for enhancing VLM safety with a visual guard module (VGM) designed to assess the harmfulness of input images for VLMs. This module endows VLMs with a dual functionality: they not only learn to generate safer responses but can also provide an interpretable harmfulness classification to justify their refusal decisions. A significant advantage of this approach is its modularity; the VGM is designed as a plug-in component, allowing for seamless integration with diverse pre-trained VLMs across various scales. Experiments show that Safe-VLM with VGM, trained on our HoliSafe, achieves state-of-the-art safety performance across multiple VLM benchmarks. Additionally, the HoliSafe-Bench itself reveals critical vulnerabilities in existing VLM models. We hope that HoliSafe and VGM will spur further research into robust and interpretable VLM safety, expanding future avenues for multimodal alignment.

Youngjin Na, Sangheon Jeong, Youngwan Lee et al.

With the growing deployment of Vision-Language Models (VLMs) in real-world applications, previously overlooked safety risks are becoming increasingly evident. In particular, seemingly innocuous multimodal inputs can combine to reveal harmful intent, leading to unsafe model outputs. While multimodal safety has received increasing attention, existing approaches often fail to address such latent risks, especially when harmfulness arises only from the interaction between modalities. We propose SIA (Safety via Intent Awareness), a training-free, intent-aware safety framework that proactively detects harmful intent in multimodal inputs and uses it to guide the generation of safe responses. SIA follows a three-stage process: (1) visual abstraction via captioning; (2) intent inference through few-shot chain-of-thought (CoT) prompting; and (3) intent-conditioned response generation. By dynamically adapting to the implicit intent inferred from an image-text pair, SIA mitigates harmful outputs without extensive retraining. Extensive experiments on safety benchmarks, including SIUO, MM-SafetyBench, and HoliSafe, show that SIA consistently improves safety and outperforms prior training-free methods.

Jeffrey Willette, Heejun Lee, Youngwan Lee et al.

The transformer's context window is vital for tasks such as few-shot learning and conditional generation as it preserves previous tokens for active memory. However, as the context lengths increase, the computational costs grow quadratically, hindering the deployment of large language models (LLMs) in real-world, long sequence scenarios. Although some recent key-value caching (KV Cache) methods offer linear inference complexity, they naively manage the stored context, prematurely evicting tokens and losing valuable information. Moreover, they lack an optimized prefill/prompt stage strategy, resulting in higher latency than even quadratic attention for realistic context sizes. In response, we introduce a novel mechanism that leverages cascading sub-cache buffers to selectively retain the most relevant tokens, enabling the model to maintain longer context histories without increasing the cache size. Our approach outperforms linear caching baselines across key benchmarks, including streaming perplexity, question answering, book summarization, and passkey retrieval, where it retains better retrieval accuracy at 1M tokens after four doublings of the cache size of 65K. Additionally, our method reduces prefill stage latency by a factor of 6.8 when compared to flash attention on 1M tokens. These innovations not only enhance the computational efficiency of LLMs but also pave the way for their effective deployment in resource-constrained environments, enabling large-scale, real-time applications with significantly reduced latency.

Heejun Lee, Geon Park, Youngwan Lee et al.

In modern large language models (LLMs), increasing the context length is crucial for improving comprehension and coherence in long-context, multi-modal, and retrieval-augmented language generation. While many recent transformer models attempt to extend their context length over a million tokens, they remain impractical due to the quadratic time and space complexities. Although recent works on linear and sparse attention mechanisms can achieve this goal, their real-world applicability is often limited by the need to re-train from scratch and significantly worse performance. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which reduces the time complexity of the attention mechanism to $O(T \log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length. We notice a pattern in the attention scores of pretrained LLMs where tokens close together tend to have similar scores, which we call ``attention locality''. Based on this observation, we utilize a novel tree-search-like algorithm that estimates the top-$k$ key tokens for a given query on the fly, which is mathematically guaranteed to have better performance than random attention pruning. In addition to improving the time complexity of the attention mechanism, we further optimize GPU memory usage by implementing KV cache offloading, which stores only $O(\log T)$ tokens on the GPU while maintaining similar decoding throughput. Experiments on benchmarks show that HiP, with its training-free nature, significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation. HiP enables pretrained LLMs to scale up to millions of tokens on commodity GPUs, potentially unlocking long-context LLM applications previously deemed infeasible.

Injoon Hwang, Haewon Park, Youngwan Lee et al.

Low-rank adaption (LoRA) is a prominent method that adds a small number of learnable parameters to the frozen pre-trained weights for parameter-efficient fine-tuning. Prompted by the question, ``Can we make its representation enough with LoRA weights solely at the final phase of finetuning without the pre-trained weights?'' In this work, we introduce Progressive Compression LoRA~(PC-LoRA), which utilizes low-rank adaptation (LoRA) to simultaneously perform model compression and fine-tuning. The PC-LoRA method gradually removes the pre-trained weights during the training process, eventually leaving only the low-rank adapters in the end. Thus, these low-rank adapters replace the whole pre-trained weights, achieving the goals of compression and fine-tuning at the same time. Empirical analysis across various models demonstrates that PC-LoRA achieves parameter and FLOPs compression rates of 94.36%/89.1% for vision models, e.g., ViT-B, and 93.42%/84.2% parameters and FLOPs compressions for language models, e.g., BERT.

Youngwan Lee, Jonghee Kim, Jeff Willette et al.

Dense computer vision tasks such as object detection and segmentation require effective multi-scale feature representation for detecting or classifying objects or regions with varying sizes. While Convolutional Neural Networks (CNNs) have been the dominant architectures for such tasks, recently introduced Vision Transformers (ViTs) aim to replace them as a backbone. Similar to CNNs, ViTs build a simple multi-stage structure (i.e., fine-to-coarse) for multi-scale representation with single-scale patches. In this work, with a different perspective from existing Transformers, we explore multi-scale patch embedding and multi-path structure, constructing the Multi-Path Vision Transformer (MPViT). MPViT embeds features of the same size~(i.e., sequence length) with patches of different scales simultaneously by using overlapping convolutional patch embedding. Tokens of different scales are then independently fed into the Transformer encoders via multiple paths and the resulting features are aggregated, enabling both fine and coarse feature representations at the same feature level. Thanks to the diverse, multi-scale feature representations, our MPViTs scaling from tiny~(5M) to base~(73M) consistently achieve superior performance over state-of-the-art Vision Transformers on ImageNet classification, object detection, instance segmentation, and semantic segmentation. These extensive results demonstrate that MPViT can serve as a versatile backbone network for various vision tasks. Code will be made publicly available at \url{https://git.io/MPViT}.

Youngwan Lee, Hyung-Il Kim, Kimin Yun et al.

Video classification researches that have recently attracted attention are the fields of temporal modeling and 3D efficient architecture. However, the temporal modeling methods are not efficient or the 3D efficient architecture is less interested in temporal modeling. For bridging the gap between them, we propose an efficient temporal modeling 3D architecture, called VoV3D, that consists of a temporal one-shot aggregation (T-OSA) module and depthwise factorized component, D(2+1)D. The T-OSA is devised to build a feature hierarchy by aggregating temporal features with different temporal receptive fields. Stacking this T-OSA enables the network itself to model short-range as well as long-range temporal relationships across frames without any external modules. Inspired by kernel factorization and channel factorization, we also design a depthwise spatiotemporal factorization module, named, D(2+1)D that decomposes a 3D depthwise convolution into two spatial and temporal depthwise convolutions for making our network more lightweight and efficient. By using the proposed temporal modeling method (T-OSA), and the efficient factorized component (D(2+1)D), we construct two types of VoV3D networks, VoV3D-M and VoV3D-L. Thanks to its efficiency and effectiveness of temporal modeling, VoV3D-L has 6x fewer model parameters and 16x less computation, surpassing a state-of-the-art temporal modeling method on both Something-Something and Kinetics-400. Furthermore, VoV3D shows better temporal modeling ability than a state-of-the-art efficient 3D architecture, X3D having comparable model capacity. We hope that VoV3D can serve as a baseline for efficient video classification.

Joong-Won Hwang, Youngwan Lee, Sungchan Oh et al.

Adversarial training deep neural networks often experience serious overfitting problem. Recently, it is explained that the overfitting happens because the sample complexity of training data is insufficient to generalize robustness. In traditional machine learning, one way to relieve overfitting from the lack of data is to use ensemble methods. However, adversarial training multiple networks is extremely expensive. Moreover, we found that there is a dilemma on choosing target model to generate adversarial examples. Optimizing attack to the members of ensemble will be suboptimal attack to the ensemble and incurs covariate shift, while attack to ensemble will weaken the members and lose the benefit from ensembling. In this paper, we propose adversarial training with Stochastic weight average (SWA); while performing adversarial training, we aggregate the temporal weight states in the trajectory of training. By adopting SWA, the benefit of ensemble can be gained without tremendous computational increment and without facing the dilemma. Moreover, we further improved SWA to be adequate to adversarial training. The empirical results on CIFAR-10, CIFAR-100 and SVHN show that our method can improve the robustness of models.

Youngwan Lee, Joong-won Hwang, Hyung-Il Kim et al.

Since many safety-critical systems, such as surgical robots and autonomous driving cars operate in unstable environments with sensor noise and incomplete data, it is desirable for object detectors to take the localization uncertainty into account. However, there are several limitations of the existing uncertainty estimation methods for anchor-based object detection. 1) They model the uncertainty of the heterogeneous object properties with different characteristics and scales, such as location (center point) and scale (width, height), which could be difficult to estimate. 2) They model box offsets as Gaussian distributions, which is not compatible with the ground truth bounding boxes that follow the Dirac delta distribution. 3) Since anchor-based methods are sensitive to anchor hyper-parameters, their localization uncertainty could also be highly sensitive to the choice of hyper-parameters. To tackle these limitations, we propose a new localization uncertainty estimation method called UAD for anchor-free object detection. Our method captures the uncertainty in four directions of box offsets (left, right, top, bottom) that are homogeneous, so that it can tell which direction is uncertain, and provide a quantitative value of uncertainty in [0, 1]. To enable such uncertainty estimation, we design a new uncertainty loss, negative power log-likelihood loss, to measure the localization uncertainty by weighting the likelihood loss by its IoU, which alleviates the model misspecification problem. Furthermore, we propose an uncertainty-aware focal loss for reflecting the estimated uncertainty to the classification score. Experimental results on COCO datasets demonstrate that our method significantly improves FCOS, by up to 1.8 points, without sacrificing computational efficiency.

Youngwan Lee, Joong-won Hwang, Sangrok Lee et al.

As DenseNet conserves intermediate features with diverse receptive fields by aggregating them with dense connection, it shows good performance on the object detection task. Although feature reuse enables DenseNet to produce strong features with a small number of model parameters and FLOPs, the detector with DenseNet backbone shows rather slow speed and low energy efficiency. We find the linearly increasing input channel by dense connection leads to heavy memory access cost, which causes computation overhead and more energy consumption. To solve the inefficiency of DenseNet, we propose an energy and computation efficient architecture called VoVNet comprised of One-Shot Aggregation (OSA). The OSA not only adopts the strength of DenseNet that represents diversified features with multi receptive fields but also overcomes the inefficiency of dense connection by aggregating all features only once in the last feature maps. To validate the effectiveness of VoVNet as a backbone network, we design both lightweight and large-scale VoVNet and apply them to one-stage and two-stage object detectors. Our VoVNet based detectors outperform DenseNet based ones with 2x faster speed and the energy consumptions are reduced by 1.6x - 4.1x. In addition to DenseNet, VoVNet also outperforms widely used ResNet backbone with faster speed and better energy efficiency. In particular, the small object detection performance has been significantly improved over DenseNet and ResNet.

Seung-Hwan Bae, Youngwan Lee, Youngjoo Jo et al.

The recent advances of convolutional detectors show impressive performance improvement for large scale object detection. However, in general, the detection performance usually decreases as the object classes to be detected increases, and it is a practically challenging problem to train a dominant model for all classes due to the limitations of detection models and datasets. In most cases, therefore, there are distinct performance differences of the modern convolutional detectors for each object class detection. In this paper, in order to build an ensemble detector for large scale object detection, we present a conceptually simple but very effective class-wise ensemble detection which is named as Rank of Experts. We first decompose an intractable problem of finding the best detections for all object classes into small subproblems of finding the best ones for each object class. We then solve the detection problem by ranking detectors in order of the average precision rate for each class, and then aggregate the responses of the top ranked detectors (i.e. experts) for class-wise ensemble detection. The main benefit of our method is easy to implement and does not require any joint training of experts for ensemble. Based on the proposed Rank of Experts, we won the 2nd place in the ILSVRC 2017 object detection competition.

Youngwan Lee, Byeonghak Yim, Huien Kim et al.

Since convolutional neural network(CNN)models emerged,several tasks in computer vision have actively deployed CNN models for feature extraction. However,the conventional CNN models have a high computational cost and require high memory capacity, which is impractical and unaffordable for commercial applications such as real-time on-road object detection on embedded boards or mobile platforms. To tackle this limitation of CNN models, this paper proposes a wide-residual-inception (WR-Inception) network, which constructs the architecture based on a residual inception unit that captures objects of various sizes on the same feature map, as well as shallower and wider layers, compared to state-of-the-art networks like ResNet. To verify the proposed networks, this paper conducted two experiments; one is a classification task on CIFAR-10/100 and the other is an on-road object detection task using a Single-Shot Multi-box Detector(SSD) on the KITTI dataset.