Deep learning has achieved remarkable success in numerous domains with help from large amounts of big data. However, the quality data labels is a concern because lack high-quality many real-world scenarios. As noisy severely degrade generalization performance deep neural networks, (robust training) becoming an important task modern applications. In this survey, we first describe problem label noise supervised perspective. Next, provide comprehensive review 62 state-of-the-art robust training...

Continual learning (CL) enables deep neural networks to adapt ever-changing data distributions. In practice, there may be scenarios where annotation is costly, leading active continual (ACL), which performs (AL) for the CL when reducing labeling cost by selecting most informative subset preferable. However, conventional AL strategies are not suitable ACL, as they focus solely on new knowledge, catastrophic forgetting of previously learned tasks. Therefore, ACL requires a strategy that can...

Noisy labels are very common in real-world training data, which lead to poor generalization on test data because of overfitting the noisy labels. In this paper, we claim that such can be avoided by "early stopping" a deep neural network before severely memorized. Then, resume early stopped using "maximal safe set," maintains collection almost certainly true-labeled samples at each epoch since stop point. Putting them all together, our novel two-phase method, called Prestopping, realizes...

Mitigating hallucinations of Large Multi-modal Models(LMMs) is crucial to enhance their reliability for general-purpose assistants. This paper shows that such LMMs can be significantly exacerbated by preceding user-system dialogues. To precisely measure this, we first present an evaluation benchmark extending popular multi-modal datasets with prepended hallucinatory dialogues generated our novel Adversarial Question Generator, which automatically generate image-related yet adversarial...

We propose a novel framework DropTop that suppresses the shortcut bias in online continual learning (OCL) while being adaptive to varying degree of incurred by continuously changing environment. By observed high-attention property bias, highly-activated features are considered candidates for debiasing. More importantly, resolving limitation environment where prior knowledge and auxiliary data not ready, two techniques---feature map fusion intensity shifting---enable us automatically...

In this paper, the measurement results on signal interference of human body communication are presented. A real is used in measurement. Also, analyzed by EM simulation. Finally, frequency band where can be without interruption proposed based and simulation

Data pruning, which aims to downsize a large training set into small informative subset, is crucial for reducing the enormous computational costs of modern deep learning. Though large-scale data collections invariably contain annotation noise and numerous robust learning methods have been developed, pruning noise-robust scenario has received little attention. With state-of-the-art Re-labeling that self-correct erroneous labels while training, it challenging identify subset induces most...

In this paper, the measurement results on effects of ground electrode according to transmission distance are presented. A biological tissue-equivalent phantom has been used in a previous but real human body is paper. Also, analyzed by EM simulation. papers K. Fujii et al. (2003)-(2002), simulation model for composed only muscle tissue. more accurate and other tissues during Finally, method which can be efficiently proposed based

Unlabeled data examples awaiting annotations contain open-set noise inevitably. A few active learning studies have attempted to deal with this for sample selection by filtering out the noisy examples. However, because focusing on purity of in a query set leads overlooking informativeness examples, best balancing and remains an important question. In paper, solve purity-informativeness dilemma learning, we propose novel Meta-Query-Net,(MQ-Net) that adaptively finds between two factors....

Recently, autoencoder (AE)-based embedding approaches have achieved state-of-the-art performance in many tasks, especially top-k recommendation with user or node classification embedding. However, we find that real-world data follow the power-law distribution respect to object sparsity. When learning AE-based embeddings of these data, dense inputs move away from sparse an space even when they are highly correlated. This phenomenon, which call polarization, obviously distorts In this paper,...

Recently, shipborne radars are being developed into fully-digital Active Electronically Scanned Array (AESA) multifunctional (MFR) that capable of electronic beam steering, simultaneously detecting and tracking multiple targets in a complex maritime environment, performing countermeasure functions real-time. However, because the development process radar is expensive owing to its high complexity, using Simulated Signal Generator (SSG) can emulate function an antenna proves effective...

동영상에서의 객체 추적은 보안, 색인 및 검색, 감시, 통신, 압축 등 다양한 분야에서 중요하다. 본 논문은 HEVC 비트스트림 상에서의 추적 방법을 제안한다. 복호화를 수행하지 않고, 상에 존재하는 움직임 벡터(MV : Motion Vector)와 부호화 크기 정보를 Spatio-Temporal Markov Random Fields (ST-MRF) 모델에 적용해 움직임의 공간적 시간적 특성을 반영한다. 변환계수를 특징점으로 활용하는 객체형태 조정 알고리즘을 ST-MRF 모델 기반 추적방법에서 나타나는 과분할에 의한 오차전파 문제를 해결한다. 제안하는 방법의 추적성능은 정확도 86.4%, 재현율 79.8%, F-measure 81.1%로 기존방법 대비 평균 F-measure는 약 0.2% 향상하지만 기존방법에서 과분할 오차전파가 두드러지는 영상에 대해서는 최대 9% 정도의 성능향상을 보인다. 전체 수행시간은 프레임 당 5.4ms이며 실시간 추적이 가능하다. Video...

In real-world continual learning scenarios, tasks often exhibit intricate and unpredictable semantic shifts, posing challenges for fixed prompt management strategies. We identify the inadequacy of universal specific prompting in handling these dynamic shifts. Universal is ineffective with abrupt changes, while struggles overfitting under mild To overcome limitations, we propose an adaptive approach that tailors minimal yet sufficient prompts based on task semantics. Our methodology,...

We propose a novel framework DropTop that suppresses the shortcut bias in online continual learning (OCL) while being adaptive to varying degree of incurred by continuously changing environment. By observed high-attention property bias, highly-activated features are considered candidates for debiasing. More importantly, resolving limitation environment where prior knowledge and auxiliary data not ready, two techniques -- feature map fusion intensity shifting enable us automatically determine...

Recently, multirotor drones have been used in various fields, and can effectively search missing people harsh environments. Infrared (IR) thermal imaging be day night to find that cannot detected visible light images. In this paper, we address the detection estimation of a person's posture dangerous environments using drone equipped with an IR camera. We detect through k-means clustering morphological operations remove false alarms based on size squareness rectangular object windows. Then,...

Learning a good distance measure for distance-based classification in time series leads to significant performance improvement many tasks. Specifically, it is critical effectively deal with variations and temporal dependencies series. However, existing metric learning approaches focus on tackling mainly using strict alignment of two sequences, thereby being not able capture dependencies. To overcome this limitation, we propose MLAT, which covers both at the same time. MLAT achieves effect as...

Catastrophic forgetting is a critical challenge in training deep neural networks. Although continual learning has been investigated as countermeasure to the problem, it often suffers from requirements of additional network components and limited scalability large number tasks. We propose novel approach by approximating true loss function using an asymmetric quadratic with one its sides overestimated. Our algorithm motivated empirical observation that parameter updates affect target functions...

온라인 상의 상품의 수가 기하 급수적으로 증가함에 따라 고객이 스스로 원하는 상품을 찾는 것이 어려워졌다. 적절한 추천은 고객의 잠재적 수요를 만족시키고 판매자의 이윤을 증대시키기에 그 중요성이 상당히 크다. 최근에는 인공신경망을 활용한 차원 축소 기법인 오토 인코더 기반의 협업 필터링 방법이 성능 면에서 두각을 나타내었다. 하지만, 인코더의 잠재 표현 분포 조정을 통해 추천 성능을 향상시키는 방법은 아직 많이 연구되지 않았다. 본 연구에서는 기반 방법에 결합되어 상품 더욱 밀집 학습방법 (DenseLR)을 제안한다. 연구의 핵심 아이디어는 유저 구매 정보 벡터들의 표현을 효과적으로 시킴에 차원에서의 효과를 강화하는 것이다. 3가지 실제 데이터 셋에 대해 기존 최첨단 연구들과 성능비교실험을 진행한 결과 제안 모든 가장 높은 보였다.

Online recommender systems should be always aligned with users' current interest to accurately suggest items that each user would like. Since usually evolves over time, the update strategy flexible quickly catch from continuously generated new user-item interactions. Existing strategies focus either on importance of interaction or learning rate for parameter, but such one-directional flexibility is insufficient adapt varying relationships between interactions and parameters. In this paper,...

Successful operation of a closed-in weapon system (CIWS) requires pre-action calibration that aligns the axis gun barrel such centroid angle outgoing friendly bullets coincides with desired direction. However, in turn an accurate estimate. This study demonstrates accuracy estimation monopulse tracking radar improves greatly by adopting frequency diversity technique.

최근에 다양한 분야에서 멀티콥터 드론을 활용하고 있으며 통해 험난한 지형의 조난자를 효과적으로 수행할 수 있다. 적외선 열화상 영상을 이용하면 주야간 환경에서 가시 영상에서 탐지할 없는 사람을 찾을 본 논문은 카메라가 탑재된 이용하여 위험한 상황에 처한 탐지하고 포즈를 추정하는 방법을 연구한다. k-means 클러스터링과 형태학적 연산을 검출하고 대상의 크기에 기반한 오류 제거를 수행한다. 그리고 검출된 피사체를 템플릿 매칭을 통하여 자세를 추정한다. 실험에서는 야간 산악에서 촬영된 서 있는 사람과 앉아있는 성공적으로 추정하였다.