Sangheum Hwang

Suho Lee, Seungwon Seo, Jihyo Kim et al.

Recently, source-free unsupervised domain adaptation (SFUDA) has emerged as a more practical and feasible approach compared to unsupervised domain adaptation (UDA) which assumes that labeled source data are always accessible. However, significant limitations associated with SFUDA approaches are often overlooked, which limits their practicality in real-world applications. These limitations include a lack of principled ways to determine optimal hyperparameters and performance degradation when the unlabeled target data fail to meet certain requirements such as a closed-set and identical label distribution to the source data. All these limitations stem from the fact that SFUDA entirely relies on unlabeled target data. We empirically demonstrate the limitations of existing SFUDA methods in real-world scenarios including out-of-distribution and label distribution shifts in target data, and verify that none of these methods can be safely applied to real-world settings. Based on our experimental results, we claim that fine-tuning a source pretrained model with a few labeled data (e.g., 1- or 3-shot) is a practical and reliable solution to circumvent the limitations of SFUDA. Contrary to common belief, we find that carefully fine-tuned models do not suffer from overfitting even when trained with only a few labeled data, and also show little change in performance due to sampling bias. Our experimental results on various domain adaptation benchmarks demonstrate that the few-shot fine-tuning approach performs comparatively under the standard SFUDA settings, and outperforms comparison methods under realistic scenarios. Our code is available at https://github.com/daintlab/fewshot-SFDA .

Jae-Hun Lee, Doyoung Yoon, ByeongMoon Ji et al.

Linear probing (LP) (and $k$-NN) on the upstream dataset with labels (e.g., ImageNet) and transfer learning (TL) to various downstream datasets are commonly employed to evaluate the quality of visual representations learned via self-supervised learning (SSL). Although existing SSL methods have shown good performances under those evaluation protocols, we observe that the performances are very sensitive to the hyperparameters involved in LP and TL. We argue that this is an undesirable behavior since truly generic representations should be easily adapted to any other visual recognition task, i.e., the learned representations should be robust to the settings of LP and TL hyperparameters. In this work, we try to figure out the cause of performance sensitivity by conducting extensive experiments with state-of-the-art SSL methods. First, we find that input normalization for LP is crucial to eliminate performance variations according to the hyperparameters. Specifically, batch normalization before feeding inputs to a linear classifier considerably improves the stability of evaluation, and also resolves inconsistency of $k$-NN and LP metrics. Second, for TL, we demonstrate that a weight decay parameter in SSL significantly affects the transferability of learned representations, which cannot be identified by LP or $k$-NN evaluations on the upstream dataset. We believe that the findings of this study will be beneficial for the community by drawing attention to the shortcomings in the current SSL evaluation schemes and underscoring the need to reconsider them.

Jihyo Kim, Jeonghyeon Kim, Sangheum Hwang

Active learning aims to identify the most informative data from an unlabeled data pool that enables a model to reach the desired accuracy rapidly. This benefits especially deep neural networks which generally require a huge number of labeled samples to achieve high performance. Most existing active learning methods have been evaluated in an ideal setting where only samples relevant to the target task, i.e., in-distribution samples, exist in an unlabeled data pool. A data pool gathered from the wild, however, is likely to include samples that are irrelevant to the target task at all and/or too ambiguous to assign a single class label even for the oracle. We argue that assuming an unlabeled data pool consisting of samples from various distributions is more realistic. In this work, we introduce new active learning benchmarks that include ambiguous, task-irrelevant out-of-distribution as well as in-distribution samples. We also propose an active learning method designed to acquire informative in-distribution samples in priority. The proposed method leverages both labeled and unlabeled data pools and selects samples from clusters on the feature space constructed via contrastive learning. Experimental results demonstrate that the proposed method requires a lower annotation budget than existing active learning methods to reach the same level of accuracy.

Jihyo Kim, Seulbi Lee, Sangheum Hwang

With the recent emergence of foundation models trained on internet-scale data and demonstrating remarkable generalization capabilities, such foundation models have become more widely adopted, leading to an expanding range of application domains. Despite this rapid proliferation, the trustworthiness of foundation models remains underexplored. Specifically, the out-of-distribution detection (OoDD) capabilities of large vision-language models (LVLMs), such as GPT-4o, which are trained on massive multi-modal data, have not been sufficiently addressed. The disparity between their demonstrated potential and practical reliability raises concerns regarding the safe and trustworthy deployment of foundation models. To address this gap, we evaluate and analyze the OoDD capabilities of various proprietary and open-source LVLMs. Our investigation contributes to a better understanding of how these foundation models represent confidence scores through their generated natural language responses. Furthermore, we propose a self-guided prompting approach, termed Reflexive Guidance (ReGuide), aimed at enhancing the OoDD capability of LVLMs by leveraging self-generated image-adaptive concept suggestions. Experimental results demonstrate that our ReGuide enhances the performance of current LVLMs in both image classification and OoDD tasks. The lists of sampled images, along with the prompts and responses for each sample are available at https://github.com/daintlab/ReGuide.

Seungwon Seo, Suho Lee, Sangheum Hwang

Utilizing large-scale pretrained models is a well-known strategy to enhance performance on various target tasks. It is typically achieved through fine-tuning pretrained models on target tasks. However, naïve fine-tuning may not fully leverage knowledge embedded in pretrained models. In this study, we introduce a novel fine-tuning method, called stochastic cross-attention (StochCA), specific to Transformer architectures. This method modifies the Transformer's self-attention mechanism to selectively utilize knowledge from pretrained models during fine-tuning. Specifically, in each block, instead of self-attention, cross-attention is performed stochastically according to the predefined probability, where keys and values are extracted from the corresponding block of a pretrained model. By doing so, queries and channel-mixing multi-layer perceptron layers of a target model are fine-tuned to target tasks to learn how to effectively exploit rich representations of pretrained models. To verify the effectiveness of StochCA, extensive experiments are conducted on benchmarks in the areas of transfer learning and domain generalization, where the exploitation of pretrained models is critical. Our experimental results show the superiority of StochCA over state-of-the-art approaches in both areas. Furthermore, we demonstrate that StochCA is complementary to existing approaches, i.e., it can be combined with them to further improve performance. Our code is available at https://github.com/daintlab/stochastic_cross_attention

Jihyo Kim, Jiin Koo, Sangheum Hwang

Deep neural networks have achieved outstanding performance over various tasks, but they have a critical issue: over-confident predictions even for completely unknown samples. Many studies have been proposed to successfully filter out these unknown samples, but they only considered narrow and specific tasks, referred to as misclassification detection, open-set recognition, or out-of-distribution detection. In this work, we argue that these tasks should be treated as fundamentally an identical problem because an ideal model should possess detection capability for all those tasks. Therefore, we introduce the unknown detection task, an integration of previous individual tasks, for a rigorous examination of the detection capability of deep neural networks on a wide spectrum of unknown samples. To this end, unified benchmark datasets on different scales were constructed and the unknown detection capabilities of existing popular methods were subject to comparison. We found that Deep Ensemble consistently outperforms the other approaches in detecting unknowns; however, all methods are only successful for a specific type of unknown. The reproducible code and benchmark datasets are available at https://github.com/daintlab/unknown-detection-benchmarks .

Kyungyul Kim, ByeongMoon Ji, Doyoung Yoon et al.

The generalization capability of deep neural networks has been substantially improved by applying a wide spectrum of regularization methods, e.g., restricting function space, injecting randomness during training, augmenting data, etc. In this work, we propose a simple yet effective regularization method named progressive self-knowledge distillation (PS-KD), which progressively distills a model's own knowledge to soften hard targets (i.e., one-hot vectors) during training. Hence, it can be interpreted within a framework of knowledge distillation as a student becomes a teacher itself. Specifically, targets are adjusted adaptively by combining the ground-truth and past predictions from the model itself. We show that PS-KD provides an effect of hard example mining by rescaling gradients according to difficulty in classifying examples. The proposed method is applicable to any supervised learning tasks with hard targets and can be easily combined with existing regularization methods to further enhance the generalization performance. Furthermore, it is confirmed that PS-KD achieves not only better accuracy, but also provides high quality of confidence estimates in terms of calibration as well as ordinal ranking. Extensive experimental results on three different tasks, image classification, object detection, and machine translation, demonstrate that our method consistently improves the performance of the state-of-the-art baselines. The code is available at https://github.com/lgcnsai/PS-KD-Pytorch.

Uichan Lee, Jeonghyeon Kim, Sangheum Hwang

Recent advances in text-to-image (T2I) diffusion models have seen rapid and widespread adoption. However, their powerful generative capabilities raise concerns about potential misuse for synthesizing harmful, private, or copyrighted content. To mitigate such risks, concept erasure techniques have emerged as a promising solution. Prior works have primarily focused on fine-tuning the denoising component (e.g., the U-Net backbone). However, recent causal tracing studies suggest that visual attribute information is localized in the early self-attention layers of the text encoder, indicating a potential alternative for concept erasing. Building on this insight, we conduct preliminary experiments and find that directly fine-tuning early layers can suppress target concepts but often degrades the generation quality of non-target concepts. To overcome this limitation, we propose High-Level Representation Misdirection (HiRM), which misdirects high-level semantic representations of target concepts in the text encoder toward designated vectors such as random directions or semantically defined directions (e.g., supercategories), while updating only early layers that contain causal states of visual attributes. Our decoupling strategy enables precise concept removal with minimal impact on unrelated concepts, as demonstrated by strong results on UnlearnCanvas and NSFW benchmarks across diverse targets (e.g., objects, styles, nudity). HiRM also preserves generative utility at low training cost, transfers to state-of-the-art architectures such as Flux without additional training, and shows synergistic effects with denoiser-based concept erasing methods.

JungWoo Chae, Jiyoon Kim, JaeWoong Choi et al.

Personalizing diffusion models using limited data presents significant challenges, including overfitting, loss of prior knowledge, and degradation of text alignment. Overfitting leads to shifts in the noise prediction distribution, disrupting the denoising trajectory and causing the model to lose semantic coherence. In this paper, we propose Adaptive Personalized Training (APT), a novel framework that mitigates overfitting by employing adaptive training strategies and regularizing the model's internal representations during fine-tuning. APT consists of three key components: (1) Adaptive Training Adjustment, which introduces an overfitting indicator to detect the degree of overfitting at each time step bin and applies adaptive data augmentation and adaptive loss weighting based on this indicator; (2)Representation Stabilization, which regularizes the mean and variance of intermediate feature maps to prevent excessive shifts in noise prediction; and (3) Attention Alignment for Prior Knowledge Preservation, which aligns the cross-attention maps of the fine-tuned model with those of the pretrained model to maintain prior knowledge and semantic coherence. Through extensive experiments, we demonstrate that APT effectively mitigates overfitting, preserves prior knowledge, and outperforms existing methods in generating high-quality, diverse images with limited reference data.

Seulbi Lee, Sangheum Hwang

Large Vision Language Models (LVLMs) have achieved remarkable progress, yet they often suffer from language bias, producing answers without relying on visual evidence. While prior work attempts to mitigate this issue through decoding strategies, architectural modifications, or curated instruction data, they typically lack a quantitative measure of how much individual training samples or tokens actually benefit from the image. In this work, we introduce Visual Information Gain (VIG), a perplexity-based metric that measures the reduction in prediction uncertainty provided by visual input. VIG enables fine-grained analysis at both sample and token levels, effectively highlighting visually grounded elements such as colors, spatial relations, and attributes. Leveraging this, we propose a VIG-guided selective training scheme that prioritizes high-VIG samples and tokens. This approach improves visual grounding and mitigates language bias, achieving superior performance with significantly reduced supervision by focusing exclusively on visually informative samples and tokens.

Sungjun Cho, Dasol Hwang, Frederic Sala et al.

Current unlearning metrics for generative models evaluate success based on reference responses or classifier outputs rather than assessing the core objective: whether the unlearned model behaves indistinguishably from a model that never saw the unwanted data. This reference-specific approach creates systematic blind spots, allowing models to appear successful while retaining unwanted knowledge accessible through alternative prompts or attacks. We address these limitations by proposing Functional Alignment for Distributional Equivalence (FADE), a novel metric that measures distributional similarity between unlearned and reference models by comparing bidirectional likelihood assignments over generated samples. Unlike existing approaches that rely on predetermined references, FADE captures functional alignment across the entire output distribution, providing a principled assessment of genuine unlearning. Our experiments on the TOFU benchmark for LLM unlearning and the UnlearnCanvas benchmark for text-to-image diffusion model unlearning reveal that methods achieving near-optimal scores on traditional metrics fail to achieve distributional equivalence, with many becoming more distant from the gold standard than before unlearning. These findings expose fundamental gaps in current evaluation practices and demonstrate that FADE provides a more robust foundation for developing and assessing truly effective unlearning methods.

JungWoo Chae, Jiyoon Kim, Sangheum Hwang

Personalizing diffusion models to specific users or concepts remains challenging, particularly when only a few reference images are available. Existing methods such as DreamBooth and Textual Inversion often overfit to limited data, causing misalignment between generated images and text prompts when attempting to balance identity fidelity with prompt adherence. While Direct Consistency Optimization (DCO) with its consistency-guided sampling partially alleviates this issue, it still struggles with complex or stylized prompts. In this paper, we propose a parallel rescaling technique for personalized diffusion models. Our approach explicitly decomposes the consistency guidance signal into parallel and orthogonal components relative to classifier free guidance (CFG). By rescaling the parallel component, we minimize disruptive interference with CFG while preserving the subject's identity. Unlike prior personalization methods, our technique does not require additional training data or expensive annotations. Extensive experiments show improved prompt alignment and visual fidelity compared to baseline methods, even on challenging stylized prompts. These findings highlight the potential of parallel rescaled guidance to yield more stable and accurate personalization for diverse user inputs.

Jeonghyeon Kim, Sangheum Hwang

Prior research on out-of-distribution detection (OoDD) has primarily focused on single-modality models. Recently, with the advent of large-scale pretrained vision-language models such as CLIP, OoDD methods utilizing such multi-modal representations through zero-shot and prompt learning strategies have emerged. However, these methods typically involve either freezing the pretrained weights or only partially tuning them, which can be suboptimal for downstream datasets. In this paper, we highlight that multi-modal fine-tuning (MMFT) can achieve notable OoDD performance. Despite some recent works demonstrating the impact of fine-tuning methods for OoDD, there remains significant potential for performance improvement. We investigate the limitation of naïve fine-tuning methods, examining why they fail to fully leverage the pretrained knowledge. Our empirical analysis suggests that this issue could stem from the modality gap within in-distribution (ID) embeddings. To address this, we propose a training objective that enhances cross-modal alignment by regularizing the distances between image and text embeddings of ID data. This adjustment helps in better utilizing pretrained textual information by aligning similar semantics from different modalities (i.e., text and image) more closely in the hyperspherical representation space. We theoretically demonstrate that the proposed regularization corresponds to the maximum likelihood estimation of an energy-based model on a hypersphere. Utilizing ImageNet-1k OoD benchmark datasets, we show that our method, combined with post-hoc OoDD approaches leveraging pretrained knowledge (e.g., NegLabel), significantly outperforms existing methods, achieving state-of-the-art OoDD performance and leading ID accuracy.

SeokHyun Seo, Jinwoo Hong, JungWoo Chae et al.

Utilizing well-trained representations in transfer learning often results in superior performance and faster convergence compared to training from scratch. However, even if such good representations are transferred, a model can easily overfit the limited training dataset and lose the valuable properties of the transferred representations. This phenomenon is more severe in ViT due to its low inductive bias. Through experimental analysis using attention maps in ViT, we observe that the rich representations deteriorate when trained on a small dataset. Motivated by this finding, we propose a novel and simple regularization method for ViT called Guided Transfer of spatial Attention (GTA). Our proposed method regularizes the self-attention maps between the source and target models. A target model can fully exploit the knowledge related to object localization properties through this explicit regularization. Our experimental results show that the proposed GTA consistently improves the accuracy across five benchmark datasets especially when the number of training data is small.

Jeongeun Park, Seungyoun Shin, Sangheum Hwang et al.

Robust learning methods aim to learn a clean target distribution from noisy and corrupted training data where a specific corruption pattern is often assumed a priori. Our proposed method can not only successfully learn the clean target distribution from a dirty dataset but also can estimate the underlying noise pattern. To this end, we leverage a mixture-of-experts model that can distinguish two different types of predictive uncertainty, aleatoric and epistemic uncertainty. We show that the ability to estimate the uncertainty plays a significant role in elucidating the corruption patterns as these two objectives are tightly intertwined. We also present a novel validation scheme for evaluating the performance of the corruption pattern estimation. Our proposed method is extensively assessed in terms of both robustness and corruption pattern estimation through a number of domains, including computer vision and natural language processing.

Jooyoung Moon, Jihyo Kim, Younghak Shin et al.

Despite the power of deep neural networks for a wide range of tasks, an overconfident prediction issue has limited their practical use in many safety-critical applications. Many recent works have been proposed to mitigate this issue, but most of them require either additional computational costs in training and/or inference phases or customized architectures to output confidence estimates separately. In this paper, we propose a method of training deep neural networks with a novel loss function, named Correctness Ranking Loss, which regularizes class probabilities explicitly to be better confidence estimates in terms of ordinal ranking according to confidence. The proposed method is easy to implement and can be applied to the existing architectures without any modification. Also, it has almost the same computational costs for training as conventional deep classifiers and outputs reliable predictions by a single inference. Extensive experimental results on classification benchmark datasets indicate that the proposed method helps networks to produce well-ranked confidence estimates. We also demonstrate that it is effective for the tasks closely related to confidence estimation, out-of-distribution detection and active learning.

Mitko Veta, Yujing J. Heng, Nikolas Stathonikos et al.

Tumor proliferation is an important biomarker indicative of the prognosis of breast cancer patients. Assessment of tumor proliferation in a clinical setting is highly subjective and labor-intensive task. Previous efforts to automate tumor proliferation assessment by image analysis only focused on mitosis detection in predefined tumor regions. However, in a real-world scenario, automatic mitosis detection should be performed in whole-slide images (WSIs) and an automatic method should be able to produce a tumor proliferation score given a WSI as input. To address this, we organized the TUmor Proliferation Assessment Challenge 2016 (TUPAC16) on prediction of tumor proliferation scores from WSIs. The challenge dataset consisted of 500 training and 321 testing breast cancer histopathology WSIs. In order to ensure fair and independent evaluation, only the ground truth for the training dataset was provided to the challenge participants. The first task of the challenge was to predict mitotic scores, i.e., to reproduce the manual method of assessing tumor proliferation by a pathologist. The second task was to predict the gene expression based PAM50 proliferation scores from the WSI. The best performing automatic method for the first task achieved a quadratic-weighted Cohen's kappa score of $κ$ = 0.567, 95% CI [0.464, 0.671] between the predicted scores and the ground truth. For the second task, the predictions of the top method had a Spearman's correlation coefficient of r = 0.617, 95% CI [0.581 0.651] with the ground truth. This was the first study that investigated tumor proliferation assessment from WSIs. The achieved results are promising given the difficulty of the tasks and weakly-labelled nature of the ground truth. However, further research is needed to improve the practical utility of image analysis methods for this task.

Sangheum Hwang, Sunggyun Park

We introduce an accurate lung segmentation model for chest radiographs based on deep convolutional neural networks. Our model is based on atrous convolutional layers to increase the field-of-view of filters efficiently. To improve segmentation performances further, we also propose a multi-stage training strategy, network-wise training, which the current stage network is fed with both input images and the outputs from pre-stage network. It is shown that this strategy has an ability to reduce falsely predicted labels and produce smooth boundaries of lung fields. We evaluate the proposed model on a common benchmark dataset, JSRT, and achieve the state-of-the-art segmentation performances with much fewer model parameters.

Kyunghyun Paeng, Sangheum Hwang, Sunggyun Park et al.

We present a unified framework to predict tumor proliferation scores from breast histopathology whole slide images. Our system offers a fully automated solution to predicting both a molecular data-based, and a mitosis counting-based tumor proliferation score. The framework integrates three modules, each fine-tuned to maximize the overall performance: An image processing component for handling whole slide images, a deep learning based mitosis detection network, and a proliferation scores prediction module. We have achieved 0.567 quadratic weighted Cohen's kappa in mitosis counting-based score prediction and 0.652 F1-score in mitosis detection. On Spearman's correlation coefficient, which evaluates predictive accuracy on the molecular data based score, the system obtained 0.6171. Our approach won first place in all of the three tasks in Tumor Proliferation Assessment Challenge 2016 which is MICCAI grand challenge.

Hyo-Eun Kim, Sangheum Hwang, Kyunghyun Cho

Latent representation learned from multi-layered neural networks via hierarchical feature abstraction enables recent success of deep learning. Under the deep learning framework, generalization performance highly depends on the learned latent representation which is obtained from an appropriate training scenario with a task-specific objective on a designed network model. In this work, we propose a novel latent space modeling method to learn better latent representation. We designed a neural network model based on the assumption that good base representation can be attained by maximizing the total correlation between the input, latent, and output variables. From the base model, we introduce a semantic noise modeling method which enables class-conditional perturbation on latent space to enhance the representational power of learned latent feature. During training, latent vector representation can be stochastically perturbed by a modeled class-conditional additive noise while maintaining its original semantic feature. It implicitly brings the effect of semantic augmentation on the latent space. The proposed model can be easily learned by back-propagation with common gradient-based optimization algorithms. Experimental results show that the proposed method helps to achieve performance benefits against various previous approaches. We also provide the empirical analyses for the proposed class-conditional perturbation process including t-SNE visualization.

Hyo-Eun Kim, Sangheum Hwang

A weakly-supervised semantic segmentation framework with a tied deconvolutional neural network is presented. Each deconvolution layer in the framework consists of unpooling and deconvolution operations. 'Unpooling' upsamples the input feature map based on unpooling switches defined by corresponding convolution layer's pooling operation. 'Deconvolution' convolves the input unpooled features by using convolutional weights tied with the corresponding convolution layer's convolution operation. The unpooling-deconvolution combination helps to eliminate less discriminative features in a feature extraction stage, since output features of the deconvolution layer are reconstructed from the most discriminative unpooled features instead of the raw one. This results in reduction of false positives in a pixel-level inference stage. All the feature maps restored from the entire deconvolution layers can constitute a rich discriminative feature set according to different abstraction levels. Those features are stacked to be selectively used for generating class-specific activation maps. Under the weak supervision (image-level labels), the proposed framework shows promising results on lesion segmentation in medical images (chest X-rays) and achieves state-of-the-art performance on the PASCAL VOC segmentation dataset in the same experimental condition.

Sangheum Hwang, Hyo-Eun Kim

Recent advances of deep learning have achieved remarkable performances in various challenging computer vision tasks. Especially in object localization, deep convolutional neural networks outperform traditional approaches based on extraction of data/task-driven features instead of hand-crafted features. Although location information of region-of-interests (ROIs) gives good prior for object localization, it requires heavy annotation efforts from human resources. Thus a weakly supervised framework for object localization is introduced. The term "weakly" means that this framework only uses image-level labeled datasets to train a network. With the help of transfer learning which adopts weight parameters of a pre-trained network, the weakly supervised learning framework for object localization performs well because the pre-trained network already has well-trained class-specific features. However, those approaches cannot be used for some applications which do not have pre-trained networks or well-localized large scale images. Medical image analysis is a representative among those applications because it is impossible to obtain such pre-trained networks. In this work, we present a "fully" weakly supervised framework for object localization ("semi"-weakly is the counterpart which uses pre-trained filters for weakly supervised localization) named as self-transfer learning (STL). It jointly optimizes both classification and localization networks simultaneously. By controlling a supervision level of the localization network, STL helps the localization network focus on correct ROIs without any types of priors. We evaluate the proposed STL framework using two medical image datasets, chest X-rays and mammograms, and achieve signiticantly better localization performance compared to previous weakly supervised approaches.