Training Strategies and Benchmarks for Weakly Supervised Multi-Label Classification

김영욱 서울대학교 대학원 2024 국내박사

Computer vision tasks in the field of deep learning are making it possible for machines to have human-level visual intelligence. Among them, multi-label image classification is a fundamental, important, and practical task in the field of computer vision that requires a holistic understanding of the scene image. However, training a deep neural network that performs multi-label classification requires expensive labeling costs. To mitigate this problem, methods for training models at low labeling costs are being actively researched. Specifically, it has been proposed to train the model with partial annotations or crowdsourced annotations instead of full annotations. However, these annotations are cheap but imperfect. Label noise included in them hinders the learning of the model and impairs the performance of the model. In this dissertation, the non-trivial and challenging problem of training multi-label classification model with these imperfect annotations is studied. First, strategies for training deep neural networks with partial annotations are focused on and studied. The key objective is to minimize the impact of label noise on model training. The learning dynamics of the model in the presence of label noise are analyzed, and a novel scheme to modify the large loss samples during training is proposed. The corruption of the model’s explanation by label noise is also analyzed, and a novel scheme to recover this damage is proposed. Experimental results show that the proposed schemes successfully minimize the influence of label noise on the model, achieving competitive performance even trained with partial annotation. In particular, on the challenging Pascal VOC dataset, a model trained with partial annotation with 20x less labeling cost than full annotation performs on par with a model trained with full annotation by applying proposed methods. The focus is secondly placed on crowdsourced annotation, and the evaluation benchmark of the existing algorithms is studied. The case of errors that omit or confuse the information registered in the database by real users is analyzed. In response to this, a novel methodology for injecting synthetic label noise that mimics the characteristics of real-world errors is proposed. This methodology can be utilized as a suitable benchmark to assess the robustness of algorithms that train models to be robust to crowdsourced label noise. Overall, it is hoped that the studies conducted in this dissertation will be helpful for training multi-label classification models in an annotation-efficient manner for ground images and remote sensing images. 딥러닝 분야에서 컴퓨터 비전 문제들은 기계가 인간 수준의 시각 지능을 갖추는 것을 가능하게 하고 있다. 그 중에서도 특히 다중 레이블 이미지 분류는 장면 이미지에 대한 전체적인 이해를 필요로 하는 컴퓨터 비전 분야의 기본적이고도 중요하며 실용적인 문제이다. 그러나 다중 레이블 분류를 수행하는 심층 신경망을 훈련시키려면 값비싼 라벨링 비용이 필요하다. 이러한 문제를 완화하기 위해 낮은 라벨링 비용으로 모델을 훈련시키는 방법이 활발히 연구되고 있다. 특히, 전체 주석 대신 부분 주석이나 크라우드소싱 주석으로 모델을 훈련시키는 방법이 제안되고 있다. 그러나 이러한 주석은 저렴한 대신 불완전하는 단점이 존재한다. 이러한 주석에 포함된 라벨 노이즈는 모델의 학습을 방해하고 모델 성능을 저하시킨다. 이 학위논문에서는 앞에서 언급한 불완전한 주석으로 다중 레이블 분류 모델을 훈련시키는 도전적인 문제를 다룬다. 먼저는 부분 주석에 초점을 맞추고 이러한 주석으로 심층 신경망을 훈련시키기 위한 전략을 다룬다. 핵심 목표는 모델 훈련에 미치는 라벨 노이즈의 영향을 최소화하는 것이다. 라벨 노이즈가 있을 때 모델의 학습 역학을 분석하고 훈련 중에 손실이 크게 발생하는 샘플을 수정 또는 조정하는 방식을 제안한다. 또한 라벨 노이즈가 모델의 예측 결과에 대한 설명을 어떻게 손상시키는지 분석하고 이러한 손상을 복구하는 방식을 제안한다. 실험 결과, 제안한 방식들은 라벨 노이즈가 모델에 미치는 영향을 성공적으로 최소화하여 부분 주석으로 학습한 경우에도 경쟁력 있는 성능을 달성하는 것으로 나타났다. 특히, 까다로운 Pascal VOC 데이터셋에서 전체 주석보다 라벨링 비용이 20배 적은 부분 주석으로 훈련된 모델이 전체 주석으로 훈련된 모델과 동등한 성능을 보였다. 두 번째로는 크라우드소싱 주석에 초점을 맞추고 기존 알고리즘을 평가하는 환경, 즉 벤치마크에 대해 다룬다. 실제 사용자가 데이터베이스에 등록하는 정보를 누락하거나 혼동하는 오류 사례를 분석한다. 이를 반영하여, 실제 오류의 특성을 모방한 인위적인 라벨 노이즈를 주입하는 방법론을 제안한다. 이 방법론은 크라우드소싱 라벨 노이즈에 강건하도록 모델을 학습시키는 알고리즘의 강건성을 평가하는데 적합한 평가 환경으로 활용될 수 있다. 본 학위논문에서 수행된 연구가 지상 이미지와 원격 탐사 이미지에 대한 다중 레이블 분류 모델을 주석 효율적으로 훈련시키는 데 있어 도움이 되기를 기대한다.

Model-based and data-driven techniques for environment-robust automatic speech recognition

강신재 서울대학교 대학원 2015 국내박사

In this thesis, we propose model-based and data-driven techniques for environment-robust automatic speech recognition. The model-based technique is the feature enhancement method in the reverberant noisy environment to improve the performance of Gaussian mixture model-hidden Markov model (HMM) system. It is based on the interacting multiple model (IMM), which was originally developed in single-channel scenario. We extend the single-channel IMM algorithm such that it can handle the multi-channel inputs under the Bayesian framework. The multi-channel IMM algorithm is capable of tracking time-varying room impulse responses and background noises by updating the relevant parameters in an on-line manner. In order to reduce the computation as the number of microphones increases, a computationally efficient algorithm is also devised. In various simulated and real environmental conditions, the performance gain of the proposed method has been confirmed. The data-driven techniques are based on deep neural network (DNN)-HMM hybrid system. In order to enhance the performance of DNN-HMM system in the adverse environments, we propose three techniques. Firstly, we propose a novel supervised pre-training technique for DNN-HMM system to achieve robust speech recognition in adverse environments. In the proposed approach, our aim is to initialize the DNN parameters such that they yield abstract features robust to acoustic environment variations. In order to achieve this, we first derive the abstract features from an early fine-tuned DNN model which is trained based on a clean speech database. By using the derived abstract features as the target values, the standard error back-propagation algorithm with the stochastic gradient descent method is performed to estimate the initial parameters of the DNN. The performance of the proposed algorithm was evaluated on Aurora-4 DB and better results were observed compared to a number of conventional pre-training methods. Secondly, a new DNN-based robust speech recognition approaches taking advantage of noise estimates are proposed. A novel part of the proposed approaches is that the time-varying noise estimates are applied to the DNN as additional inputs. For this, we extract the noise estimates in a frame-by-frame manner from the IMM algorithm which has been known to show good performance in tracking slowly-varying background noise. The performance of the proposed approaches is evaluated on Aurora-4 DB and better performance is observed compared to the conventional DNN-based robust speech recognition algorithms. Finally, a new approach to DNN-based robust speech recognition using soft target labels is proposed. The soft target labeling means that each target value of the DNN output is not restricted to 0 or 1 but takes non negative values in (0,1) and their sum equals 1. In this study, the soft target labels are obtained from the forward-backward algorithm well-known in HMM training. The proposed method makes the DNN training be more robust in noisy and unseen conditions. The performance of the proposed approach was evaluated on Aurora-4 DB and various mismatched noise test conditions, and found better compared to the conventional hard target labeling method. Furthermore, in the data-driven approaches, an integrated technique using above three algorithms and model-based technique is described. In matched and mismatched noise conditions, the performance results are discussed. In matched noise conditions, the initialization method for the DNN was effective to enhance the recognition performance. In mismatched noise conditions, the combination of using the noise estimates as an DNN input and soft target labels showed the best recognition results in all the tested combinations of the proposed techniques.

기계학습의 정확도 향상을 위한 레이블 노이즈 제거 알고리즘

무하마드 암마르 말릭 조선대학교 대학원 2017 국내석사

머신러닝을 위한 분류기 학습 데이터에서 각 데이터의 클래스가 항상 정확할 수 없기 때문에 기계학습데이터 레이블링에 오류가 포함될 가능성이 높다. 예를 들어 의학 자동 진단 분야에 서, 질병의 분류 및 진단에 대한 오류가 포함될 가능성이 항상 존재한다. 기계학습 알고리즘 은 입력 데이터의 클래스 레이블링 정확도에 많은 영향을 받기 때문에, 분류기의 성능은 잠재 적인 오류들이 포함된 데이터들에 의해 결정이 된다. 본 논문에서는 기계학습 데이터에 오류 가 존재할 때, 이 오류를 인지하고, 제거하는 알고리즘을 제시한다. 이러한 오류 데이터들의 대부분은 기계학습에서 사용되는 분류기에 의해 명확하게 구분되지 않는 구간에 대부분 존재 한다는 것에 착안하여, 기계학습에 가장 많이 사용되는 SVM 분류기를 기준으로 학습데이터 의 유클리안 위치를 이용하여 오류가 포함되었을 가능성이 높은 데이터를 인지하는 방법과, 이와 반대로 SVM분류기에서 멀리 떨어져, 오류가 발생하지 않았을 가능성이 높은 데이터를 활용하여 오류가 포함되었을 가능성이 높은 데이터를 다시 레이블링하는 두가지 종류의 알고 리즘을 제시하였다. 제안된 방법들 여러 가지 종류의 데이터를 이용하여 효율적으로 레이블 링 에러를 제거할 수 있다는 것을 검증하였다. Performance of machine learning classifiers is heavily dependent on labeling quality of datasets. Generally, human supervision is required for the labeling of instances in datasets. This labeling can be erroneous, and detecting such erroneous examples from the dataset is extremely important. In this work we discuss some of the machine learning approaches to deal with the problem of label noise in datasets. The experiments are conducted on some of the widely used datasets in the machine learning community. Firstly, a clustering based technique for relabeling of instances in datasets is studied. Secondly, a similarity based technique that utilizes the concept of Euclidean distance for cleaning of label noise. The instances having similar scores with positive and negative classes are selected for expert review. Lastly, an improved majority filter is proposed. Our experiments show that the improved majority filter is faster as compared to the conventional majority filter. We also compare the performance of proposed method with majority and consensus filter in terms of precision, recall and F_1 Score.

Light and Chemistry at the Interface of Theory and Experiment

Ulcickas, James RW Purdue University ProQuest Dissertations & Theses 2020 해외박사(DDOD)

Optics are a powerful probe of chemical structure that can often be linked to theoretical predictions, providing robustness as a measurement tool. Not only do optical interactions like second harmonic generation (SHG), single and two-photon excited fluorescence (TPEF), and infrared absorption provide chemical specificity at the molecular and macromolecular scale, but the ability to image enables mapping heterogeneous behavior across complex systems such as biological tissue. This thesis will discuss nonlinear and linear optics, leveraging theoretical predictions to provide frameworks for interpreting analytical measurement. In turn, the causal mechanistic understanding provided by these frameworks will enable structurally specific quantitative tools with a special emphasis on application in biological imaging. The thesis will begin with an introduction to 2nd order nonlinear optics and the polarization analysis thereof, covering both the Jones framework for polarization analysis and the design of experiment. Novel experimental architectures aimed at reducing 1/f noise in polarization analysis will be discussed, leveraging both rapid modulation in time through electro-optic modulators (Chapter 2), as well as fixed-optic spatial modulation approaches (Chapter 3). In addition, challenges in polarization-dependent imaging within turbid systems will be addressed with the discussion of a theoretical framework to model SHG occurring from unpolarized light (Chapter 4). The application of this framework to thick tissue imaging for analysis of collagen local structure can provide a method for characterizing changes in tissue morphology associated with some common cancers (Chapter 5). In addition to discussion of nonlinear optical phenomena, a novel mechanism for electric dipole allowed fluorescence-detected circular dichroism will be introduced (Chapter 6). Tackling challenges associated with label-free chemically specific imaging, the construction of a novel infrared hyperspectral microscope for chemical classification in complex mixtures will be presented (Chapter 7). The thesis will conclude with a discussion of the inherent disadvantages in taking the traditional paradigm of modeling and measuring chemistry separately and provide the multi-agent consensus equilibrium (MACE) framework as an alternative to the classic meet-in-the-middle approach (Chapter 8). Spanning topics from pure theoretical descriptions of light-matter interaction to full experimental work, this thesis aims to unify these two fronts.

작업장소음의 발생원대책 강화를 위한 저소음기계 인증제도 개선방안 연구

조영준 전북대학교 환경대학원 2019 국내석사

Noisy equipment is pointed out as a major cause of noise complaints and deterioration of the working environment by increasing noise pollution, whether outdoors or indoors. An effective model for noise control is the source, path, and receiver model. And the noise control at the source is known as the most economical and efficient countermeasure. In advanced countries, various policies are being enforced to reduce equipment noise level at it’s source. In Korea, both bligatory and recommended rules for enlarging the use of low noise equipment are enforced together. However, the use ratio of low noise equipment is still low at actual industrial sites. This study analyses the problems of domestic noise policies related to noisy equipment, and presents a plan for improving the implementation of expansion policy of low noise equipment based on this analysis. This paper proposes that the equipment satisfying the compulsory rule be able to use the environmental certification mark, the standards of which are defined in recommended rules, to improve the linkage between the two policy measures. This thesis also suggests expanding the range of equipment to be supervised and using new noise label that has grades and shows detailed noise level of the equipment to improve environmental certification system. In addition, supplementing the construction permission system and supporting international certification system are offered to expand the use of low noise equipment. Keyword : Noise labeling, Noise control, Low noise equipment certification system 높은 소음을 배출하는 기계류는 공장과 공사장의 소음도를 증가시켜 소음민원과 작업환경 악화를 초래하는 주요 원인으로 지적되고 있다. 고소음 배출 기계에 대한 소음방지 대책은 크게 발생원 대책, 전파경로 대책, 수음자 대책으로 나눌 수 있으며, 발생원 대책이 가장 경제적이고 효율적인 대책으로 알려져 있다. 국내?외에서는 고소음 기계에 대한 발생원 대책을 강화하기 위하여 다양한 정책들을 시행하고 있다. 국내에서도 저소음 기계의 사용을 확대하기 위한 의무규정과 권고규정이 함께 시행되고 있다. 그러나 실제 산업현장에서 저소음기계의 활용도는 낮은 편이다. 본 연구에서는 고소음 발생기계 관련 국내 소음정책의 문제점을 분석하고 이를 토대로 저소음 기계 사용 확대 정책의 개선방안을 제시하였다. 의무규정과 권고규정이 중복되는 관리항목에 대해서는 의무규정에서 정한 소음규정을 따르도록 하고, 의무규정을 만족하는 제품에 대해서는 권고규정에서 정한 환경표지를 부착할 수 있도록 하여 두 제도간에 연계성을 높이는 방안을 제안하였다. 또한, 관리대상을 확대하고, 인증기준보다 높은 수준의 저소음 기계에 대해서는 2~3등급으로 차등화하여 소음도를 표시하도록 환경표지 도안을 개선하는 방안을 제안하였다. 아울러 적합성평가기관의 국내 도입을 통해 국제인증을 지원하고, 허가제도를 통해 저소음기계의 사용을 확대하는 방안을 함께 제안하였다. 핵심단어 : 저소음기계, 소음표지, 소음도검사제도, 저소음기계인증제도

Regularization using Noise samples Identified by the Feature norm for Face recognition

김태성 아주대학교 일반대학원 2024 국내석사

Face recognition is a task that involves comparing two images of a face and determining whether they belong to the same person. Face recognition can be applied in a variety of environments, including surveillance systems. However, the performance of the deep learning model used for face recognition can be affected by the quality of the image. Therefore, recent studies on face recognition using deep learning have suggested taking image quality into consideration. Some studies have used feature norms, which is the L2 norm of extracted features from images using a deep learning model, to measure the image quality. However, previous studies have lacked analysis of why the feature norms correspond to image quality. This thesis presents a new hypothesis that a higher sample's feature norms indicate that the samples are similar to other samples learned by the deep learning model. We also demonstrate that this hypothesis can be used to distinguish noise samples. Additionally, we introduce a new regularization technique that uses noise samples to improve face recognition performance in low-resolution environments.

노이즈 데이터 정제를 통한 분류모델 성능 향상

정운국 국민대학교 소프트웨어융합대학원 2023 국내석사

Probing Malat1 in Collective Cancer Invasion

Zhu, Ninghao The Pennsylvania State University ProQuest Dissert 2022 해외박사(DDOD)

Unveiling the mechanisms of cancer invasion and predicting the progression of cancer invasion are two vital directions to mitigate the healthcare burden from cancer invasion. Cancer cells invade collectively with leader-follower organization. However, how leader cells are regulated during the dynamic invasion process remains poorly understood. Using a FRET nanobiosensor that tracks long noncoding RNA (lncRNA) dynamics in live single cells, we monitored the spatiotemporal distribution of lncRNA during collective cancer invasion. We show that lncRNA MALAT1 is dynamically regulated in the invading fronts of cancer cells and patient-derived organoids. The abundance, diffusivity, and distribution of MALAT1 transcripts are distinct between leader and follower cells. MALAT1 expression increases when a cell acquires the leader cell role and decreases when the migration process stops. Transient knockdown of MALAT1 prevents the formation of leader cells and abolishes the migration of cancer cells. Combining the MALAT1 biosensing with a chimeric invasion assay, we analyzed the invasiveness of transurethral resection of bladder tumor samples. The human dissociated bladder tumor cells of various cancer stages show different levels of dissemination after self-assembly on the CIA surface, correlating with the MALAT1 expression. Taken together, this work provides a promising biosensing platform for bladder cancer prognosis.