Junha Song

Chaoning Zhang, Chenshuang Zhang, Junha Song et al.

Masked autoencoders are scalable vision learners, as the title of MAE \cite{he2022masked}, which suggests that self-supervised learning (SSL) in vision might undertake a similar trajectory as in NLP. Specifically, generative pretext tasks with the masked prediction (e.g., BERT) have become a de facto standard SSL practice in NLP. By contrast, early attempts at generative methods in vision have been buried by their discriminative counterparts (like contrastive learning); however, the success of mask image modeling has revived the masking autoencoder (often termed denoising autoencoder in the past). As a milestone to bridge the gap with BERT in NLP, masked autoencoder has attracted unprecedented attention for SSL in vision and beyond. This work conducts a comprehensive survey of masked autoencoders to shed insight on a promising direction of SSL. As the first to review SSL with masked autoencoders, this work focuses on its application in vision by discussing its historical developments, recent progress, and implications for diverse applications.

Junha Song, Jungsoo Lee, In So Kweon et al.

This paper presents a simple yet effective approach that improves continual test-time adaptation (TTA) in a memory-efficient manner. TTA may primarily be conducted on edge devices with limited memory, so reducing memory is crucial but has been overlooked in previous TTA studies. In addition, long-term adaptation often leads to catastrophic forgetting and error accumulation, which hinders applying TTA in real-world deployments. Our approach consists of two components to address these issues. First, we present lightweight meta networks that can adapt the frozen original networks to the target domain. This novel architecture minimizes memory consumption by decreasing the size of intermediate activations required for backpropagation. Second, our novel self-distilled regularization controls the output of the meta networks not to deviate significantly from the output of the frozen original networks, thereby preserving well-trained knowledge from the source domain. Without additional memory, this regularization prevents error accumulation and catastrophic forgetting, resulting in stable performance even in long-term test-time adaptation. We demonstrate that our simple yet effective strategy outperforms other state-of-the-art methods on various benchmarks for image classification and semantic segmentation tasks. Notably, our proposed method with ResNet-50 and WideResNet-40 takes 86% and 80% less memory than the recent state-of-the-art method, CoTTA.

Junha Song, Kwanyong Park, InKyu Shin et al.

Prior to the deployment of robotic systems, pre-training the deep-recognition models on all potential visual cases is infeasible in practice. Hence, test-time adaptation (TTA) allows the model to adapt itself to novel environments and improve its performance during test time (i.e., lifelong adaptation). Several works for TTA have shown promising adaptation performances in continuously changing environments. However, our investigation reveals that existing methods are vulnerable to dynamic distributional changes and often lead to overfitting of TTA models. To address this problem, this paper first presents a robust TTA framework with compound domain knowledge management. Our framework helps the TTA model to harvest the knowledge of multiple representative domains (i.e., compound domain) and conduct the TTA based on the compound domain knowledge. In addition, to prevent overfitting of the TTA model, we devise novel regularization which modulates the adaptation rates using domain-similarity between the source and the current target domain. With the synergy of the proposed framework and regularization, we achieve consistent performance improvements in diverse TTA scenarios, especially on dynamic domain shifts. We demonstrate the generality of proposals via extensive experiments including image classification on ImageNet-C and semantic segmentation on GTA5, C-driving, and corrupted Cityscapes datasets.

Junha Song, Tae Soo Kim, Junha Kim et al.

This paper aims to adapt the source model to the target environment, leveraging small user feedback (i.e., labeled target data) readily available in real-world applications. We find that existing semi-supervised domain adaptation (SemiSDA) methods often suffer from poorly improved adaptation performance when directly utilizing such feedback data, as shown in Figure 1. We analyze this phenomenon via a novel concept called Negatively Biased Feedback (NBF), which stems from the observation that user feedback is more likely for data points where the model produces incorrect predictions. To leverage this feedback while avoiding the issue, we propose a scalable adapting approach, Retrieval Latent Defending. This approach helps existing SemiSDA methods to adapt the model with a balanced supervised signal by utilizing latent defending samples throughout the adaptation process. We demonstrate the problem caused by NBF and the efficacy of our approach across various benchmarks, including image classification, semantic segmentation, and a real-world medical imaging application. Our extensive experiments reveal that integrating our approach with multiple state-of-the-art SemiSDA methods leads to significant performance improvements.

Junha Song, Byeongho Heo, Geonmo Gu et al.

When humans describe a visual scene, they do not process the entire image uniformly; instead, they selectively fixate on regions relevant to their intended description. In contrast, current multimodal large language models (MLLMs) attend to all visual tokens at each generation step, leading to diluted focus and unnecessary computational overhead. In this work, we introduce Gaze Attention, a novel mechanism that enables MLLMs to selectively attend to task-relevant visual regions during generation. Specifically, we spatially group visual embeddings-stored as key-value caches-into compact gaze regions, each represented by a lightweight descriptor. At each decoding step, the model dynamically selects the most relevant regions and restricts attention to them, reducing redundant computation while enhancing focus. To mitigate the loss of global context caused by localized attention, we further propose learnable context tokens appended to each image or frame, allowing the model to maintain holistic visual awareness. Extensive experiments on image and video understanding benchmarks demonstrate that Gaze Attention matches or surpasses dense-attention baselines, while using up to 90% fewer visual KV entries in the attention computation.

Junha Song, Sangdoo Yun, Dongyoon Han et al.

A dominant assumption in Multimodal Language Model (MLLM) research is that its performance is largely inherited from the LLM backbone, given its immense parameter scale and remarkable capabilities. This has created a void in the understanding of the vision encoder, which determines how MLLMs perceive images. The recent shift in MLLM training paradigms, from Supervised Finetuning (SFT) to Reinforcement Learning (RL), magnifies this oversight-namely, the significant lack of analysis on how such training reshapes the vision encoder as well as the MLLM. To address this, we first investigate the impact of training strategies on MLLMs, where RL shows a clear advantage over SFT in strongly vision-related VQA benchmarks. Motivated by this, we conduct a critical yet under-explored analysis of the vision encoder of MLLMs through diverse and in-depth experiments, ranging from ImageNet classification and segmentation to gradient visualization. Our results demonstrate that MLLM's post-training strategy (i.e., SFT or RL) not only leads to distinct outcomes on MLLM downstream tasks, but also fundamentally reshapes MLLM's underlying visual representations. Specifically, the key finding of our study is that RL produces stronger and precisely localized visual representations compared to SFT, boosting the ability of the vision encoder for MLLM. We then reframe our findings into a simple recipe for building strong vision encoders for MLLMs, Preference-Instructed Vision OpTimization (PIVOT). When integrated into MLLMs, a PIVOT-trained vision encoder outperforms even larger and more heavily-trained counterparts, despite requiring less than 1% of the computational cost of standard vision pretraining. This result opens an effective and efficient path for advancing the vision backbones of MLLMs. Project page available at https://june-page.github.io/pivot/

Junha Song, Yongsik Jo, So Yeon Min et al. · cmu

Image captioning is fundamental for applications like video-grounded chatbot systems and navigation robots, yet deploying such models on local devices is challenging due to the high computational demands of multimodal LLMs (MLLMs). To address this, we first build lightweight captioning models using a 125M-parameter language model, 56 times smaller than LLaMA-7B, and evaluate their performance not only on single-sentence but on detailed captioning tasks. We obtain surprising results showing that our model can achieve performance comparable to MLLMs, suggesting its potential to serve as a strong captioning specialist for on-device applications. While promising, our model also exhibits a limitation: like other MLLMs, it suffers from occasional captioning errors. We investigate the underlying causes and observe that the problems stem from ineffective attention mechanisms and limited visual representations. To alleviate them, we develop a novel captioning framework, Sharp-Eyed Refinement, which enhances caption quality by refining coarse descriptions into more precise captions. At its core, DeepLens improves visual grounding by re-examining the informative regions identified in the initial glance. Experimental results demonstrate the superiority of our model over both recent lightweight captioning methods and MLLMs in detailed captioning and even in long-range video QA tasks.