arXiv subjects

Yusung Kim

Publications and source records attributed to Yusung Kim.

13 recordsLinked to original sources

arXiv2026

Learning Generalizable Skill Policy with Data-Efficient Unsupervised RL

Unsupervised Reinforcement Learning (URL) aims to pre-train scalable, skill-conditioned policies without extrinsic rewards, serving as a foundation for downstream control tasks. Despite recent progress, we argue that current off-policy URL methods are limited by two critical, overlooked bottlenecks: (1) non-stationary skill semantics and (2) brittle generalization. To address these challenges, we propose GenDa (Generalizable Data-efficient Agent), a unified framework for robust unsupervised reinforcement learning. First, we introduce a skill relabeling mechanism to mitigate non-stationarity and significantly improve data efficiency for pre-training. Second, we propose a Complementary Information Bottleneck (CIB), encouraging the learned skill policy to focus on ego-centric features and become robust to distribution shifts for downstream tasks. Through various experiments, we demonstrate that GenDa significantly enhances the scalability of URL with superior generalizability and data efficiency. Our code and videos are available at https://ihatebroccoli.github.io/official-GenDa.

cs.LG↗

arXiv2026

Dosimetric Impact of Hidden Input Parameters in Inverse Optimization Algorithms for GYN HDR Brachytherapy

Inverse optimization (IO) algorithms are used in GYN HDR brachytherapy planning, with user parameter settings embedded in commercial TPS. To examine the dosimetric influence of hidden input parameters in three IO algorithms-IPSA, HIPO, and MCO-for GYN HDR brachytherapy across two applicator types. In-house implementations of IPSA, HIPO, and MCO were implemented and evaluated against retrospectively generated commercial TPS plans (Oncentra Brachy) using identical clinical input parameters across 24 cervical cancer cases (18 T&O; 6 T&O+Needles (T&O+N)). Each IO algorithm was assessed using 1k combinations of hidden parameters (e.g., dwell-time modulation constraints, convergence thresholds). Cumulative DVH curves and dosimetric indices (HR-CTV D98/D90, OAR D2cc) were compared with commercial plans. Standard deviations (SD) of DVH differences were used to characterize sensitivity to hidden parameters. For HR-CTV, SD values in T&O+N cases reached 23.0 Gy and 7.1 Gy for MCO and HIPO, respectively, with corresponding average values of 55.8 Gy and 19.7 Gy. In T&O cases, HR-CTV SD values reached 4.9 Gy and 3.3 Gy for HIPO and IPSA, respectively, with average values of 20.1 Gy and 8.6 Gy. MCO exhibited the highest sensitivity, followed by HIPO and IPSA. T&O+N cases showed greater sensitivity than T&O cases. Absolute differences in HR-CTV D90 (D98) relative to commercial algorithms reached up to 33.3 Gy (28.4) for T&O+N cases and 10.8 Gy (8.5) for T&O cases. For OARs, absolute D2cc differences in T&O+N (T&O) cases reached up to 8.6 Gy (2.3) for rectum, 17 Gy (10.2) for bladder, 14.8 Gy (3.9) for sigmoid, and 7.0 Gy (8.1) for bowel. Hidden input parameter settings significantly impact on GYN HDR plans, with target coverage up to 28.4 Gy across IO algorithms for both T&O and T&O+N cases. The findings in this study shown the potential to improve plans through hidden input parameter optimization.

physics.med-ph↗

arXiv2025

Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning

Existing offline hierarchical reinforcement learning methods rely on high-level policy learning to generate subgoal sequences. However, their efficiency degrades as task horizons increase, and they lack effective strategies for stitching useful state transitions across different trajectories. We propose Graph-Assisted Stitching (GAS), a novel framework that formulates subgoal selection as a graph search problem rather than learning an explicit high-level policy. By embedding states into a Temporal Distance Representation (TDR) space, GAS clusters semantically similar states from different trajectories into unified graph nodes, enabling efficient transition stitching. A shortest-path algorithm is then applied to select subgoal sequences within the graph, while a low-level policy learns to reach the subgoals. To improve graph quality, we introduce the Temporal Efficiency (TE) metric, which filters out noisy or inefficient transition states, significantly enhancing task performance. GAS outperforms prior offline HRL methods across locomotion, navigation, and manipulation tasks. Notably, in the most stitching-critical task, it achieves a score of 88.3, dramatically surpassing the previous state-of-the-art score of 1.0. Our source code is available at: https://github.com/qortmdgh4141/GAS.

cs.LG↗

arXiv2025

Self-Predictive Dynamics for Generalization of Vision-based Reinforcement Learning

Vision-based reinforcement learning requires efficient and robust representations of image-based observations, especially when the images contain distracting (task-irrelevant) elements such as shadows, clouds, and light. It becomes more important if those distractions are not exposed during training. We design a Self-Predictive Dynamics (SPD) method to extract task-relevant features efficiently, even in unseen observations after training. SPD uses weak and strong augmentations in parallel, and learns representations by predicting inverse and forward transitions across the two-way augmented versions. In a set of MuJoCo visual control tasks and an autonomous driving task (CARLA), SPD outperforms previous studies in complex observations, and significantly improves the generalization performance for unseen observations. Our code is available at https://github.com/unigary/SPD.

cs.CV↗

arXiv2025

Dream to Generalize: Zero-Shot Model-Based Reinforcement Learning for Unseen Visual Distractions

Model-based reinforcement learning (MBRL) has been used to efficiently solve vision-based control tasks in highdimensional image observations. Although recent MBRL algorithms perform well in trained observations, they fail when faced with visual distractions in observations. These task-irrelevant distractions (e.g., clouds, shadows, and light) may be constantly present in real-world scenarios. In this study, we propose a novel self-supervised method, Dream to Generalize (Dr. G), for zero-shot MBRL. Dr. G trains its encoder and world model with dual contrastive learning which efficiently captures task-relevant features among multi-view data augmentations. We also introduce a recurrent state inverse dynamics model that helps the world model to better understand the temporal structure. The proposed methods can enhance the robustness of the world model against visual distractions. To evaluate the generalization performance, we first train Dr. G on simple backgrounds and then test it on complex natural video backgrounds in the DeepMind Control suite, and the randomizing environments in Robosuite. Dr. G yields a performance improvement of 117% and 14% over prior works, respectively. Our code is open-sourced and available at https://github.com/JeongsooHa/DrG.git

cs.CV↗

arXiv2025

Self-supervised One-Stage Learning for RF-based Multi-Person Pose Estimation

In the field of Multi-Person Pose Estimation (MPPE), Radio Frequency (RF)-based methods can operate effectively regardless of lighting conditions and obscured line-of-sight situations. Existing RF-based MPPE methods typically involve either 1) converting RF signals into heatmap images through complex preprocessing, or 2) applying a deep embedding network directly to raw RF signals. The first approach, while delivering decent performance, is computationally intensive and time-consuming. The second method, though simpler in preprocessing, results in lower MPPE accuracy and generalization performance. This paper proposes an efficient and lightweight one-stage MPPE model based on raw RF signals. By sub-grouping RF signals and embedding them using a shared single-layer CNN followed by multi-head attention, this model outperforms previous methods that embed all signals at once through a large and deep CNN. Additionally, we propose a new self-supervised learning (SSL) method that takes inputs from both one unmasked subgroup and the remaining masked subgroups to predict the latent representations of the masked data. Empirical results demonstrate that our model improves MPPE accuracy by up to 15 in PCKh@0.5 compared to previous methods using raw RF signals. Especially, the proposed SSL method has shown to significantly enhance performance improvements when placed in new locations or in front of obstacles at RF antennas, contributing to greater performance gains as the number of people increases. Our code and dataset is open at Github. https://github.com/sshnan7/SOSPE .

cs.CV↗

arXiv2024

Cross-Domain Semantic Segmentation on Inconsistent Taxonomy using VLMs

The challenge of semantic segmentation in Unsupervised Domain Adaptation (UDA) emerges not only from domain shifts between source and target images but also from discrepancies in class taxonomies across domains. Traditional UDA research assumes consistent taxonomy between the source and target domains, thereby limiting their ability to recognize and adapt to the taxonomy of the target domain. This paper introduces a novel approach, Cross-Domain Semantic Segmentation on Inconsistent Taxonomy using Vision Language Models (CSI), which effectively performs domain-adaptive semantic segmentation even in situations of source-target class mismatches. CSI leverages the semantic generalization potential of Visual Language Models (VLMs) to create synergy with previous UDA methods. It leverages segment reasoning obtained through traditional UDA methods, combined with the rich semantic knowledge embedded in VLMs, to relabel new classes in the target domain. This approach allows for effective adaptation to extended taxonomies without requiring any ground truth label for the target domain. Our method has shown to be effective across various benchmarks in situations of inconsistent taxonomy settings (coarse-to-fine taxonomy and open taxonomy) and demonstrates consistent synergy effects when integrated with previous state-of-the-art UDA methods. The implementation is available at http://github.com/jkee58/CSI.

cs.CV↗

arXiv2021

Optimal Needle Placement for Prostate Rotating-Shield Brachytherapy (RSBT)

Purpose: To present an efficient NEEdle Position Optimization (NEEPO) algorithm for prostate rotating shield brachytherapy (RSBT). With RSBT, the increased flexibility beyond conventional high-dose-rate brachytherapy (HDR-BT) due to the partially shielded radiation source has been shown by Adams et al. in 2020 to enable improved urethra sparing (23.1%), enhanced dose escalation (29.9%), or both, with 20 needles without NEEPO-optimized positions. Within this regime of improved dosimetry, we propose in this work that the benefits of RSBT can be maintained while also reducing the number of needles needed for the delivery. The goal of NEEPO is to provide the capability to further increase the dosimetric benefit of RSBT and to minimize the number of needles needed to satisfy a dosimetric goal. Methods: The NEEPO algorithm generates a needle pool for a given patient and then iteratively constructs a subset of needles from the pool based on relative needle importance as determined by total dwell times within needles. The NEEPO algorithm is based on a convex optimization formulation using a quadratic dosimetric penalty function, dwell time regularization by total variation, and a block sparsity regularization term to enable iterative removal of low-importance needles. RSBT treatment plans for 26 patients were generated using single fraction prescriptions with both dose escalation and urethra sparing goals, and compared to baseline HDR-BT treatment plans.

physics.med-ph↗

arXiv2019

Deep segmentation networks predict survival of non-small cell lung cancer

Non-small-cell lung cancer (NSCLC) represents approximately 80-85% of lung cancer diagnoses and is the leading cause of cancer-related death worldwide. Recent studies indicate that image-based radiomics features from positron emission tomography-computed tomography (PET/CT) images have predictive power on NSCLC outcomes. To this end, easily calculated functional features such as the maximum and the mean of standard uptake value (SUV) and total lesion glycolysis (TLG) are most commonly used for NSCLC prognostication, but their prognostic value remains controversial. Meanwhile, convolutional neural networks (CNN) are rapidly emerging as a new premise for cancer image analysis, with significantly enhanced predictive power compared to other hand-crafted radiomics features. Here we show that CNN trained to perform the tumor segmentation task, with no other information than physician contours, identify a rich set of survival-related image features with remarkable prognostic value. In a retrospective study on 96 NSCLC patients before stereotactic-body radiotherapy (SBRT), we found that the CNN segmentation algorithm (U-Net) trained for tumor segmentation in PET/CT images, contained features having strong correlation with 2- and 5-year overall and disease-specific survivals. The U-net algorithm has not seen any other clinical information (e.g. survival, age, smoking history) than the images and the corresponding tumor contours provided by physicians. Furthermore, through visualization of the U-Net, we also found convincing evidence that the regions of progression appear to match with the regions where the U-Net features identified patterns that predicted higher likelihood of death. We anticipate our findings will be a starting point for more sophisticated non-intrusive patient specific cancer prognosis determination.

eess.IV↗

arXiv2017

Magnetic Tunnel Junction Mimics Stochastic Cortical Spiking Neurons

Brain-inspired computing architectures attempt to mimic the computations performed in the neurons and the synapses in the human brain in order to achieve its efficiency in learning and cognitive tasks. In this work, we demonstrate the mapping of the probabilistic spiking nature of pyramidal neurons in the cortex to the stochastic switching behavior of a Magnetic Tunnel Junction in presence of thermal noise. We present results to illustrate the efficiency of neuromorphic systems based on such probabilistic neurons for pattern recognition tasks in presence of lateral inhibition and homeostasis. Such stochastic MTJ neurons can also potentially provide a direct mapping to the probabilistic computing elements in Belief Networks for performing regenerative tasks.

cs.ET↗

arXiv2017

Fast dose optimization for rotating shield brachytherapy

Purpose: To provide a fast computational method, based on the proximal graph solver (POGS) - a convex optimization solver using the alternating direction method of multipliers (ADMM), for calculating an optimal treatment plan in rotating shield brachytherapy (RSBT). RSBT treatment planning has more degrees of freedom than conventional high-dose-rate brachytherapy (HDR-BT) due to the addition of emission direction, and this necessitates a fast optimization technique to enable clinical usage. // Methods: The multi-helix RSBT (H-RSBT) delivery technique was considered with five representative cervical cancer patients. Treatment plans were generated for all patients using the POGS method and the previously considered commercial solver IBM CPLEX. The rectum, bladder, sigmoid, high-risk clinical target volume (HR-CTV), and HR-CTV boundary were the structures considered in our optimization problem, called the asymmetric dose-volume optimization with smoothness control. Dose calculation resolution was 1x1x3 mm^3 for all cases. The H-RSBT applicator has 6 helices, with 33.3 mm of translation along the applicator per helical rotation and 1.7 mm spacing between dwell positions, yielding 17.5 degree emission angle spacing per 5 mm along the applicator.// Results: For each patient, HR-CTV D90, HR-CTV D100, rectum D2cc, sigmoid D2cc, and bladder D2cc matched within 1% for CPLEX and POGS. Also, we obtained similar EQD2 figures between CPLEX and POGS. POGS was around 18 times faster than CPLEX. Over all patients, total optimization times were 32.1-65.4 seconds for CPLEX and 2.1-3.9 seconds for POGS. // Conclusions: POGS substantially reduced treatment plan optimization time around 18 times for RSBT with similar HR-CTV D90, OAR D2cc values, and EQD2 figure relative to CPLEX, which is significant progress toward clinical translation of RSBT. POGS is also applicable to conventional HDR-BT.

physics.med-ph↗

arXiv2014

Spin Orbit Torque Based Electronic Neuron

A device based on current-induced spin-orbit torque (SOT) that functions as an electronic neuron is proposed in this work. The SOT device implements an artificial neuron's thresholding (transfer) function. In the first step of a two-step switching scheme, a charge current places the magnetization of a nano-magnet along the hard-axis i.e. an unstable point for the magnet. In the second step, the SOT device (neuron) receives a current (from the synapses) which moves the magnetization from the unstable point to one of the two stable states. The polarity of the synaptic current encodes the excitatory and inhibitory nature of the neuron input, and determines the final orientation of the magnetization. A resistive crossbar array, functioning as synapses, generates a bipolar current that is a weighted sum of the inputs. The simulation of a two layer feed-forward Artificial Neural Network (ANN) based on the SOT electronic neuron shows that it consumes ~3X lower power than a 45nm digital CMOS implementation, while reaching ~80% accuracy in the classification of one hundred images of handwritten digits from the MNIST dataset.

cs.ET↗

arXiv2013

DSTT-MRAM: Differential Spin Hall MRAM for On-chip Memories

A new device structure for spin transfer torque based magnetic random access memory is proposed for on-chip memory applications. Our device structure exploits spin Hall effect to create a differential memory cell that exhibits fast and energy-efficient write operation. Moreover, due to inherently differential device structure, fast and reliable read operation can be performed. Our simulation study shows 10X improvement in write energy over the standard 1T1R STT-MRAM memory cell, and 1.6X faster read operation compared to single-ended sensing (as in standard 1T1R STT-MRAMs). The bit-cell characteristics are promising for high performance on-chip memory applications.

cond-mat.mes-hall↗