arXiv subjects

Seungho Lee

Publications and source records attributed to Seungho Lee.

29 records · Page 2Linked to original sources

arXiv2021

Evaluation for Weakly Supervised Object Localization: Protocol, Metrics, and Datasets

Weakly-supervised object localization (WSOL) has gained popularity over the last years for its promise to train localization models with only image-level labels. Since the seminal WSOL work of class activation mapping (CAM), the field has focused on how to expand the attention regions to cover objects more broadly and localize them better. However, these strategies rely on full localization supervision for validating hyperparameters and model selection, which is in principle prohibited under the WSOL setup. In this paper, we argue that WSOL task is ill-posed with only image-level labels, and propose a new evaluation protocol where full supervision is limited to only a small held-out set not overlapping with the test set. We observe that, under our protocol, the five most recent WSOL methods have not made a major improvement over the CAM baseline. Moreover, we report that existing WSOL methods have not reached the few-shot learning baseline, where the full-supervision at validation time is used for model training instead. Based on our findings, we discuss some future directions for WSOL.

cs.CV↗

arXiv2021

Laser astrophysics experiment on the amplification of magnetic fields by shock-induced interfacial instabilities

Laser experiments are becoming established as a new tool for astronomical research that complements observations and theoretical modeling. Localized strong magnetic fields have been observed at a shock front of supernova explosions. Experimental confirmation and identification of the physical mechanism for this observation are of great importance in understanding the evolution of the interstellar medium. However, it has been challenging to treat the interaction between hydrodynamic instabilities and an ambient magnetic field in the laboratory. Here, we developed an experimental platform to examine magnetized Richtmyer-Meshkov instability (RMI). The measured growth velocity was consistent with the linear theory, and the magnetic-field amplification was correlated with RMI growth. Our experiment validated the turbulent amplification of magnetic fields associated with the shock-induced interfacial instability in astrophysical conditions for the first time. Experimental elucidation of fundamental processes in magnetized plasmas is generally essential in various situations such as fusion plasmas and planetary sciences.

physics.plasm-ph↗

arXiv2021

Railroad is not a Train: Saliency as Pseudo-pixel Supervision for Weakly Supervised Semantic Segmentation

Existing studies in weakly-supervised semantic segmentation (WSSS) using image-level weak supervision have several limitations: sparse object coverage, inaccurate object boundaries, and co-occurring pixels from non-target objects. To overcome these challenges, we propose a novel framework, namely Explicit Pseudo-pixel Supervision (EPS), which learns from pixel-level feedback by combining two weak supervisions; the image-level label provides the object identity via the localization map and the saliency map from the off-the-shelf saliency detection model offers rich boundaries. We devise a joint training strategy to fully utilize the complementary relationship between both information. Our method can obtain accurate object boundaries and discard co-occurring pixels, thereby significantly improving the quality of pseudo-masks. Experimental results show that the proposed method remarkably outperforms existing methods by resolving key challenges of WSSS and achieves the new state-of-the-art performance on both PASCAL VOC 2012 and MS COCO 2014 datasets.

cs.CV↗

arXiv2021

Orbital angular momentum and current-induced motion of a Skyrmion-textured domain wall in a ferromagnetic nanotube

We theoretically study the current-induced dynamics of a domain wall in a ferromagnetic nanotube by developing a theory for the orbital angular momentum of a domain wall and the current-induced torque on it. Specifically, a domain wall with nontrivial magnetization winding along the circumference is shown to possess finite orbital angular momentum, which is proportional to the product of its Skyrmion charge and position, and the current is shown to exert a torque changing the orbital angular momentum of the domain wall and thereby drives it. The current-induced torque is interpreted as the transfer of orbital angular momentum from electrons to the domain wall, which occurs due to the emergent magnetic field associated with the Skyrmion charge. Our results reveal a hitherto unrecognized utility of the orbital degree of freedom of magnetic solitons.

cond-mat.mes-hall↗

arXiv2020

Evaluating Weakly Supervised Object Localization Methods Right

Weakly-supervised object localization (WSOL) has gained popularity over the last years for its promise to train localization models with only image-level labels. Since the seminal WSOL work of class activation mapping (CAM), the field has focused on how to expand the attention regions to cover objects more broadly and localize them better. However, these strategies rely on full localization supervision to validate hyperparameters and for model selection, which is in principle prohibited under the WSOL setup. In this paper, we argue that WSOL task is ill-posed with only image-level labels, and propose a new evaluation protocol where full supervision is limited to only a small held-out set not overlapping with the test set. We observe that, under our protocol, the five most recent WSOL methods have not made a major improvement over the CAM baseline. Moreover, we report that existing WSOL methods have not reached the few-shot learning baseline, where the full-supervision at validation time is used for model training instead. Based on our findings, we discuss some future directions for WSOL.

cs.CV↗

arXiv2019

Resource Optimized Neural Architecture Search for 3D Medical Image Segmentation

Neural Architecture Search (NAS), a framework which automates the task of designing neural networks, has recently been actively studied in the field of deep learning. However, there are only a few NAS methods suitable for 3D medical image segmentation. Medical 3D images are generally very large; thus it is difficult to apply previous NAS methods due to their GPU computational burden and long training time. We propose the resource-optimized neural architecture search method which can be applied to 3D medical segmentation tasks in a short training time (1.39 days for 1GB dataset) using a small amount of computation power (one RTX 2080Ti, 10.8GB GPU memory). Excellent performance can also be achieved without retraining(fine-tuning) which is essential in most NAS methods. These advantages can be achieved by using a reinforcement learning-based controller with parameter sharing and focusing on the optimal search space configuration of macro search rather than micro search. Our experiments demonstrate that the proposed NAS method outperforms manually designed networks with state-of-the-art performance in 3D medical image segmentation.

eess.IV↗

arXiv2019

Peta-Pascal Pressure Driven by Fast Isochoric Heating with Multi-Picosecond Intense Laser Pulse

Fast isochoric laser heating is a scheme to heat a matter with relativistic-intensity ($>$ 10$^{18}$ W/cm$^2$) laser pulse or X-ray free electron laser pulse. The fast isochoric laser heating has been studied for creating efficiently ultra-high-energy-density (UHED) state. We demonstrate an fast isochoric heating of an imploded dense plasma using a multi-picosecond kJ-class petawatt laser with an assistance of externally applied kilo-tesla magnetic fields for guiding fast electrons to the dense plasma.The UHED state with 2.2 Peta-Pascal is achieved experimentally with 4.6 kJ of total laser energy that is one order of magnitude lower than the energy used in the conventional implosion scheme. A two-dimensional particle-in-cell simulation reveals that diffusive heating from a laser-plasma interaction zone to the dense plasma plays an essential role to the efficient creation of the UHED state.

physics.plasm-ph↗

arXiv2019

Cooperative Multi-Agent Reinforcement Learning Framework for Scalping Trading

We explore deep Reinforcement Learning(RL) algorithms for scalping trading and knew that there is no appropriate trading gym and agent examples. Thus we propose gym and agent like Open AI gym in finance. Not only that, we introduce new RL framework based on our hybrid algorithm which leverages between supervised learning and RL algorithm and uses meaningful observations such order book and settlement data from experience watching scalpers trading. That is very crucial information for traders behavior to be decided. To feed these data into our model, we use spatio-temporal convolution layer, called Conv3D for order book data and temporal CNN, called Conv1D for settlement data. Those are preprocessed by episode filter we developed. Agent consists of four sub agents divided to clarify their own goal to make best decision. Also, we adopted value and policy based algorithm to our framework. With these features, we could make agent mimic scalpers as much as possible. In many fields, RL algorithm has already begun to transcend human capabilities in many domains. This approach could be a starting point to beat human in the financial stock market, too and be a good reference for anyone who wants to design RL algorithm in real world domain. Finally, weexperiment our framework and gave you experiment progress.

cs.AI↗

arXiv2018

Dynamic Multivariate Functional Data Modeling via Sparse Subspace Learning

Multivariate functional data from a complex system are naturally high-dimensional and have complex cross-correlation structure. The complexity of data structure can be observed as that (1) some functions are strongly correlated with similar features, while some others may have almost no cross-correlations with quite diverse features; and (2) the cross-correlation structure may also change over time due to the system evolution. With this regard, this paper presents a dynamic subspace learning method for multivariate functional data modeling. In particular, we consider different functions come from different subspaces, and only functions of the same subspace have cross-correlations with each other. The subspaces can be automatically formulated and learned by reformatting the problem as a sparse regression. By allowing but regularizing the regression change over time, we can describe the cross-correlation dynamics. The model can be efficiently estimated by the fast iterative shrinkage-thresholding algorithm (FISTA), and the features of every subspace can be extracted using the smooth multi-channel functional PCA. Numerical studies together with case studies demonstrate the efficiency and applicability of the proposed methodology.

stat.ML↗

arXiv2018

Super-ponderomotive electron acceleration in blowout plasma heated by multi-picosecond relativistic intensity laser pulse

The dependence of the mean kinetic energy of laser-accelerated electrons on the laser intensity, so-called ponderomotive scaling, was derived theoretically with consideration of the motion of a single electron in oscillating laser fields. This scaling explains well the experimental results obtained with high-intensity pulses and durations shorter than a picosecond; however, this scaling is no longer applicable to the multi-picosecond (multi-ps) facility experiments. Here, we experimentally clarified the generation of the super-ponderomotive-relativistic electrons (SP-REs) through multi-ps relativistic laser-plasma interactions using prepulse-free LFEX laser pulses that were realized using a plasma mirror (PM). The SP-REs are produced with direct laser acceleration assisted by the self-generated quasi-static electric field and with loop-injected direct acceleration by the self- generated quasi-static magnetic field, which grow in a blowout plasma heated by a multi-ps laser pulse. Finally, we theoretically derive the threshold pulse duration to boost the acceleration of REs, which provides an important insight into the determination of laser pulse duration at kilojoule- petawatt laser facilities.

physics.plasm-ph↗

arXiv2017

Magnetized Fast Isochoric Laser Heating for Efficient Creation of Ultra-High-Energy-Density States

The quest for the inertial confinement fusion (ICF) ignition is a grand challenge, as exemplified by extraordinary large laser facilities. Fast isochoric heating of a pre-compressed plasma core with a high-intensity short-pulse laser is an attractive and alternative approach to create ultra-high-energy-density states like those found in ICF ignition sparks. This avoids the ignition quench caused by the hot spark mixing with the surrounding cold fuel, which is the crucial problem of the currently pursued ignition scheme. High-intensity lasers efficiently produce relativistic electron beams (REB). A part of the REB kinetic energy is deposited in the core, and then the heated region becomes the hot spark to trigger the ignition. However, only a small portion of the REB collides with the core because of its large divergence. Here we have demonstrated enhanced laser-to-core energy coupling with the magnetized fast isochoric heating. The method employs a kilo-tesla-level magnetic field that is applied to the transport region from the REB generation point to the core which results in guiding the REB along the magnetic field lines to the core. 7.7 $\pm$ 1.3 % of the maximum coupling was achieved even with a relatively small radial area density core ($ρR$ $\sim$ 0.1 g/cm$^2$). The guided REB transport was clearly visualized in a pre-compressed core by using Cu-$K_α$ imaging technique. A simplified model coupled with the comprehensive diagnostics yields 6.2\% of the coupling that agrees fairly with the measured coupling. This model also reveals that an ignition-scale areal density core ($ρR$ $\sim$ 0.4 g/cm$^2$) leads to much higher laser-to-core coupling ($>$ 15%), this is much higher than that achieved by the current scheme.

physics.plasm-ph↗