Semi-supervised learning (SSL) uses unlabeled data to improve the performance of machine learning models when labeled data is scarce. However, its real-world applications often face the label distribution mismatch problem, in which the unlabeled dataset includes instances whose ground-truth labels are absent from the labeled training dataset. Recent studies, referred to as safe SSL, have addressed this issue by using both classification and out-of-distribution (OOD) detection. However, the existing methods may suffer from overconfidence in deep neural networks, leading to increased SSL errors because of high confidence in incorrect pseudo-labels or OOD detection. To address this, we propose a novel method, CaliMatch, which calibrates both the classifier and the OOD detector to foster safe SSL. CaliMatch presents adaptive label smoothing and temperature scaling, which eliminates the need to manually tune the smoothing degree for effective calibration. We give a theoretical justification for why improving the calibration of both the classifier and the OOD detector is crucial in safe SSL. Extensive evaluations on CIFAR-10, CIFAR-100, SVHN, TinyImageNet, and ImageNet demonstrate that CaliMatch outperforms the existing methods in safe SSL tasks.

Distribution shifts between training and test data undermine the reliability of deep neural networks, challenging real-world applications across domains and subpopulations. While distributionally robust optimization (DRO) methods like GroupDRO aim to improve robustness by optimizing worst-case performance over predefined groups, their use of a single global classifier can be restrictive when facing substantial inter-environment variability. We propose Multi-Expert Distributionally Robust Optimization (MEDRO), a novel extension of GroupDRO designed to address such complex shifts. MEDRO employs a shared feature extractor with $m$ environment-specific expert classifier heads, and introduces a min-max objective over all $m^{2}$ expert-environment pairings, explicitly modeling cross-environment risks. This expanded uncertainty set captures fine-grained distributional variations that a single classifier might overlook. Empirical evaluations on a range of standard distribution shift benchmarks demonstrate that MEDRO often achieves robust predictive performance compared to existing methods. Furthermore, MEDRO offers practical inference strategies, such as ensembling or gating mechanisms, for typical scenarios where environment labels are unavailable at test time. Our findings suggest MEDRO as a promising step toward resilient and generalizable machine learning under real-world distribution shifts.

Photoplethysmography (PPG) signals from wearable devices have expanded the accessibility of heart rate estimation. Recent advances in deep learning have significantly improved the generalizability of heart rate estimation from PPG signals. However, these models exhibit performance degradation when used for new subjects with different PPG distributions. Although previous studies have attempted subject-specific training and fine-tuning techniques, they require labeled data for each new subject, limiting their practicality. In response, we explore the application of domain adaptation techniques using only unlabeled PPG signals from the target subject. However, naive domain adaptation approaches do not adequately account for the variability in PPG signals among different subjects in the training dataset. Furthermore, they overlook the possibility that the heart rate range of the target subject may only partially overlap with that of the source subjects. To address these limitations, we propose a novel multi-source partial domain adaptation method, GAussian-based dUaL-level weighting (GAUL), designed for the PPG-based heart rate estimation, formulated as a regression task. GAUL considers and adjusts the contribution of relevant source data at the domain and sample levels during domain adaptation. The experimental results on three benchmark datasets demonstrate that our method outperforms existing domain adaptation approaches, enhancing the heart rate estimation accuracy for new subjects without requiring additional labeled data. The code is available at: https://github.com/Im-JihyunKim/GAUL .

As the scale of high-performance computing systems (HPC) continues to expand, the frequency of failures within the system also increases. The follow-up steps in the event of a failure are necessary because efficient prevention measures are absent. These actions, however, result in decreased operating efficiency as the system normalizes. To solve this problem, recent research has utilized deep learning to predict failures in advance by examining log messages. One common method uses log ID extracted from log messages as input data. However, these methods could lead to information loss in log messages and may not accurately capture log message properties. In this study, we propose the prior failure detection in large-scale systems using language models (FALL) that reflects the properties of log messages. The FALL uses log messages as text data and only uses normal data for training. The FALL makes use of sharpening, which concentrates on the minimal lexical change between normal and abnormal log messages. In addition, by leveraging the characteristics of log messages with limited vocabularies, the FALL utilizes only some tokens for anomaly detection. The experiments' results from log data from real industry-based HPC systems show that the FALL achieves superior performance in early failure detection.

In pixel-based deep reinforcement learning (DRL), learning representations of states that change because of an agent's action or interaction with the environment poses a critical challenge in improving data efficiency. Recent data-efficient DRL studies have integrated DRL with self-supervised learning (SSL) and data augmentation to learn state representations from given interactions. However, some methods have difficulties in explicitly capturing evolving state representations or in selecting data augmentations for appropriate reward signals. Our goal is to explicitly learn the inherent dynamics that change with an agent's intervention and interaction with the environment. We propose masked and inverse dynamics modeling (MIND), which uses masking augmentation and fewer hyperparameters to learn agent-controllable representations in changing states. Our method is comprised of a self-supervised multitask learning that leverages a transformer architecture, which captures the spatiotemporal information underlying in the highly correlated consecutive frames. MIND uses two tasks to perform self-supervised multitask learning: masked modeling and inverse dynamics modeling. Masked modeling learns the static visual representation required for control in the state, and inverse dynamics modeling learns the rapidly evolving state representation with agent intervention. By integrating inverse dynamics modeling as a complementary component to masked modeling, our method effectively learns evolving state representations. We evaluate our method by using discrete and continuous control environments with limited interactions. MIND outperforms previous methods across benchmarks and significantly improves data efficiency. The code is available at https://github.com/dudwojae/MIND.

Text-guided diffusion models have become a popular tool in image synthesis, known for producing high-quality and diverse images. However, their application to editing real images often encounters hurdles primarily due to the text condition deteriorating the reconstruction quality and subsequently affecting editing fidelity. Null-text Inversion (NTI) has made strides in this area, but it fails to capture spatial context and requires computationally intensive per-timestep optimization. Addressing these challenges, we present Noise Map Guidance (NMG), an inversion method rich in a spatial context, tailored for real-image editing. Significantly, NMG achieves this without necessitating optimization, yet preserves the editing quality. Our empirical investigations highlight NMG's adaptability across various editing techniques and its robustness to variants of DDIM inversions.

Addressing the limitations of text as a source of accurate layout representation in text-conditional diffusion models, many works incorporate additional signals to condition certain attributes within a generated image. Although successful, previous works do not account for the specific localization of said attributes extended into the three dimensional plane. In this context, we present a conditional diffusion model that integrates control over three-dimensional object placement with disentangled representations of global stylistic semantics from multiple exemplar images. Specifically, we first introduce depth disentanglement training to leverage the relative depth of objects as an estimator, allowing the model to identify the absolute positions of unseen objects through the use of synthetic image triplets. We also introduce soft guidance, a method for imposing global semantics onto targeted regions without the use of any additional localization cues. Our integrated framework, Compose and Conquer (CnC), unifies these techniques to localize multiple conditions in a disentangled manner. We demonstrate that our approach allows perception of objects at varying depths while offering a versatile framework for composing localized objects with different global semantics.

Classification of multisensor signals is an important problem in maintaining stable process operations, particularly in advancing predictive modeling for early detection of abnormal states. Self-supervised learning methods, one of the representation learnings, have been widely studied. However, they have focused on using unlabeled data. In this study, we aim to address the challenge of effectively utilizing fully labeled data for modeling multisensor signals. We introduce supervised contrastive learning (SCL) for the classification of multisensor signals. Our training framework involves a two-step process: SCL for encoder pretraining with time-series data augmentations, and classifier training with the pretrained encoder. Our method exhibits superior performance, outperforming traditional supervised learning approaches by a substantial margin. Furthermore, we demonstrate the practical applicability of our approach for early prediction problems through experiments conducted with real-process data obtained from automobile engine manufacturing. Our work offers a promising method for multisensor signal analysis and early fault detection in manufacturing industries.

Cerebral palsy (CP) is a neurological disorder that severely limits physical movements in everyday life. For individuals with CP, precise movement patterns are critical to improving their physical abilities. However, because of the time and expense involved, obtaining feedback from qualified rehabilitation professionals could be difficult. To address this problem, we propose a multimodal human action recognition (HAR) framework that integrates time-series data on skeletons and angles for upper body exercise for individuals with CP. Previous, studies typically used skeleton data as input to HAR. However, given the repetitive motion sequence, angles also are a crucial factor. In this study, we create video data of upper body exercises. Using the data, we estimate the skeleton using BlazePose, a human pose estimation (HPE) algorithm, and calculate the critical angles from the estimated skeletons. Our proposed framework incorporates graph-based models for skeletons and recurrent-based models for angles with an attention mechanism, allowing identification of important temporal information. It is shown that our framework performs better than unimodal HAR models. By reducing the costs associated with exercise feedback, we expect our approach to ultimately improve the physical health of individuals with CP.

Class imbalance problems have been reported as a major issue in various applications. Classification becomes further complicated when an imbalance occurs in time series data sets. To address time series data, it is necessary to consider their characteristics (i.e., high dimensionality, high correlations, and multimodality). Oversampling is a well-known approach for addressing this problem; however, such an approach does not appropriately consider the characteristics of time series data. This paper addresses these limitations by presenting a model-based oversampling approach, a boundary-focused generative adversarial network (BFGAN). The proposed BFGAN employs a specifically designed additional label for reflecting the importance of a sample's position in data space. Furthermore, the BFGAN generates artificial samples after taking into consideration a sample's multimodality and importance by using a suitable modified GAN structure. We present empirical results that reveal a significant improvement in the quality of the generated data when the proposed BFGAN is used as an oversampling algorithm for an imbalanced multimodal time series data set.

Root cause analysis (RCA) methods for effectively identifying the critical causes of abnormal processes have attracted attention because manufacturing processes have grown in scale and complexity. However, the existing methods for building automatic RCA models suffer from the disadvantage of typically requiring expert knowledge. In addition, without a dataset representing the causal relationship of multivariate processes, it is difficult to provide useful information for RCA. Although data-driven RCA methods have been proposed, most are based on classification models. Given that product quality is defined as a continuous variable in many manufacturing industries, classification models are limited in deriving root causes affecting the product quality level. In this article, we propose a regression model-based RCA method, which we call quality-discriminative localization, consisting of a convolutional neural network (CNN)-based activation mapping of multisensor signal data. In our proposed method, the CNN predicts the product quality of a continuous variable. Activation mapping then extracts causal maps that highlight significant sensor signals for each product. To identify the root causes, we generate a root cause map from the weighted sum of quality and causal maps. We consider root causes as locations of abnormal processes and processing times from localized activation scores on the root cause map. We experimentally demonstrate the usefulness of the proposed method with simulated data and real process data from a steel manufacturing process. Our results show that the proposed method successfully identifies root causes with distinct sensor signal patterns.

We present a hybrid algorithm called parallel simulated annealing with a greedy algorithm (PSAGA) to learn Bayesian network structures. This work focuses on simulated annealing and its parallelization with memoization to accelerate the search process. At each step of the local search, a hybrid search method combining simulated annealing with a greedy algorithm was adopted. The proposed PSAGA aims to achieve both the efficiency of parallel search and the effectiveness of a more exhaustive search. The Bayesian Dirichlet equivalence metric was used to determine an optimal structure for PSAGA. The proposed PSAGA was evaluated on seven well-known Bayesian network benchmarks generated at random. We first conducted experiments to evaluate the computational time performance of the proposed parallel search. We then compared PSAGA with existing variants of simulated annealing-based algorithms to evaluate the quality of the learned structure. Overall, the experimental results demonstrate that the proposed PSAGA shows better performance than the alternatives in terms of computational time and accuracy.

Abstract Bayesian networks (BNs) have been widely used in causal analysis because they can express the statistical relationship between significant variables. To gain superior causal analysis results, numerous studies have emphasized the importance of combining a knowledge‐based approach and a data‐based approach. However, combining these two approaches is a difficult task because it can reduce the effectiveness of the BN structure learning. Further, the learning schemes of BNs for computational efficiency can cause an inadequate causal analysis. To address these problems, we propose a knowledge‐driven BN structure calibration algorithm for rich causal semantics. We first present an algorithm that can efficiently identify the subnetworks that can be altered to satisfy the learning condition of the BNs. We then reflect experts' knowledge to reduce erroneous causalities from the learned network. Experiments on various simulation and benchmark data sets were conducted to examine the properties of the proposed method and to compare its performance with an existing method. Further, an experimental study with real data from semiconductor fabrication plants demonstrated that the proposed method provided superior performance in improving structural accuracy.

Hotelling's T2control chart is widely used as a representative method to efficiently monitor multivariate processes. However, they have some parametric restrictions that may not be applicable for modern manufacturing systems complicated. In the present study we propose a clustering algorithm-based control chart that overcomes the limitation posed by the parametric assumption in existing control chart methods. The simulation results showed that the proposed clustering algorithm-based control charts outperformed Hotelling's T2control charts especially when process data follow the nonnormal distributions.

Various analyses of the complex behavior in ambient air pollutants have been conducted to extract their implicit patterns and meaningful information. In the present study, we conducted some statistical analyses to identify daily, seasonal, and spatial patterns of particulate matters (PM10) in Seoul, Korea. We used the daily PM10mass concentration data observed at 25 different monitoring sites in Seoul, Korea from 2005 to 2009. Analysis of variance and a k-means clustering algorithm were used to investigate seasonal and spatial patterns of PM10concentrations. Moreover, we used a bootstrap method to calculate the probabilities that PM10concentrations exceeded the environment limit or the comprehensive air quality index in different months.

Isolating the variable or set of variables responsible for an out-of-control signal is a challenging task in multivariate statistical process control. Several fault isolation approaches have been proposed. However, all assumed a multivariate normal distribution on the process data, an assumption that limits their applicability in many situations. In the present study we propose a nonparametric fault isolation approach based on a hybrid novelty score (HNS). A simulation study was conducted to examine the performance of our proposed HNS-based fault isolation approach, and its results were compared with both parametric and nonparametric T2decomposition approaches. The performance of our approach was superior in the simulation to either parametric or nonparametric T2decompositions. This was especially true in nonnormal situations.

A frontier-based tree-pruning algorithm (FBP) is proposed. The new method has an order of computational complexity comparable to cost-complexity pruning (CCP). Regarding tree pruning, it provides a full spectrum of information: namely, (1) given the value of the penalization parameter λ, it gives the decision tree specified by the complexity-penalization approach; (2) given the size of a decision tree, it provides the range of the penalization parameter λ, within which the complexity-penalization approach renders this tree size; (3) it finds the tree sizes that are inadmissible—no matter what the value of the penalty parameter is, the resulting tree based on a complexity-penalization framework will never have these sizes. Simulations on real data sets reveal a “surprise:” in the complexity-penalization approach, most of the tree sizes are inadmissible. FBP facilitates a more faithful implementation of cross validation (CV), which is favored by simulations. Using FBP, a stability analysis of CV is proposed.