Feng Zhou

Junliang Lyu, Yixuan Zhang, Xiaoling Lu et al.

This work addresses a key limitation in current federated learning approaches, which predominantly focus on homogeneous tasks, neglecting the task diversity on local devices. We propose a principled integration of multi-task learning using multi-output Gaussian processes (MOGP) at the local level and federated learning at the global level. MOGP handles correlated classification and regression tasks, offering a Bayesian non-parametric approach that naturally quantifies uncertainty. The central server aggregates the posteriors from local devices, updating a global MOGP prior redistributed for training local models until convergence. Challenges in performing posterior inference on local devices are addressed through the Pólya-Gamma augmentation technique and mean-field variational inference, enhancing computational efficiency and convergence rate. Experimental results on both synthetic and real data demonstrate superior predictive performance, OOD detection, uncertainty calibration and convergence rate, highlighting the method's potential in diverse applications. Our code is publicly available at https://github.com/JunliangLv/task_diversity_BFL.

Yixuan Zhang, Jiabin Luo, Zhenggang Wang et al.

Fairness concerns are increasingly critical as machine learning models are deployed in high-stakes applications. While existing fairness-aware methods typically intervene at the model level, they often suffer from high computational costs, limited scalability, and poor generalization. To address these challenges, we propose a Bayesian data selection framework that ensures fairness by aligning group-specific posterior distributions of model parameters and sample weights with a shared central distribution. Our framework supports flexible alignment via various distributional discrepancy measures, including Wasserstein distance, maximum mean discrepancy, and $f$-divergence, allowing geometry-aware control without imposing explicit fairness constraints. This data-centric approach mitigates group-specific biases in training data and improves fairness in downstream tasks, with theoretical guarantees. Experiments on benchmark datasets show that our method consistently outperforms existing data selection and model-based fairness methods in both fairness and accuracy.

Yixuan Zhang, Boyu Li, Zenan Ling et al.

Discrimination can occur when the underlying unbiased labels are overwritten by an agent with potential bias, resulting in biased datasets that unfairly harm specific groups and cause classifiers to inherit these biases. In this paper, we demonstrate that despite only having access to the biased labels, it is possible to eliminate bias by filtering the fairest instances within the framework of confident learning. In the context of confident learning, low self-confidence usually indicates potential label errors; however, this is not always the case. Instances, particularly those from underrepresented groups, might exhibit low confidence scores for reasons other than labeling errors. To address this limitation, our approach employs truncation of the confidence score and extends the confidence interval of the probabilistic threshold. Additionally, we incorporate with co-teaching paradigm for providing a more robust and reliable selection of fair instances and effectively mitigating the adverse effects of biased labels. Through extensive experimentation and evaluation of various datasets, we demonstrate the efficacy of our approach in promoting fairness and reducing the impact of label bias in machine learning models.

Quyu Kong, Yixuan Zhang, Yang Liu et al.

Temporal Point Processes (TPPs) have been widely used for event sequence modeling, but they often struggle to incorporate rich textual event descriptions effectively. Conversely, while Large Language Models (LLMs) have been shown remarkable capabilities in processing textual data, they lack mechanisms for handling temporal dynamics. To bridge this gap, we introduce Language-TPP, a unified framework that integrates TPPs with LLMs for enhanced event sequence modeling. Language-TPP introduces a novel temporal encoding mechanism that converts continuous time intervals into specialized byte-tokens, enabling seamless integration with standard LLM architectures. This approach allows Language-TPP to achieve state-of-the-art performance across multiple TPP tasks, including event time prediction, type prediction, and intensity estimation, on five datasets. Additionally, we demonstrate that incorporating temporal information significantly improves the quality of generated event descriptions.

Yixuan Zhang, Zhidong Li, Yang Wang et al.

Machine learning algorithms often struggle to eliminate inherent data biases, particularly those arising from unreliable labels, which poses a significant challenge in ensuring fairness. Existing fairness techniques that address label bias typically involve modifying models and intervening in the training process, but these lack flexibility for large-scale datasets. To address this limitation, we introduce a data selection method designed to efficiently and flexibly mitigate label bias, tailored to more practical needs. Our approach utilizes a zero-shot predictor as a proxy model that simulates training on a clean holdout set. This strategy, supported by peer predictions, ensures the fairness of the proxy model and eliminates the need for an additional holdout set, which is a common requirement in previous methods. Without altering the classifier's architecture, our modality-agnostic method effectively selects appropriate training data and has proven efficient and effective in handling label bias and improving fairness across diverse datasets in experimental evaluations.

Yixun Zhang, Lizhi Wang, Junjun Zhao et al.

Camera-based object detection systems play a vital role in autonomous driving, yet they remain vulnerable to adversarial threats in real-world environments. Existing 2D and 3D physical attacks, due to their focus on texture optimization, often struggle to balance physical realism and attack robustness. In this work, we propose 3D Gaussian-based Adversarial Attack (3DGAA), a novel adversarial object generation framework that leverages the full 14-dimensional parameterization of 3D Gaussian Splatting (3DGS) to jointly optimize geometry and appearance in physically realizable ways. Unlike prior works that rely on patches or texture optimization, 3DGAA jointly perturbs both geometric attributes (shape, scale, rotation) and appearance attributes (color, opacity) to produce physically realistic and transferable adversarial objects. We further introduce a physical filtering module that filters outliers to preserve geometric fidelity, and a physical augmentation module that simulates complex physical scenarios to enhance attack generalization under real-world conditions. We evaluate 3DGAA on both virtual benchmarks and physical-world setups using miniature vehicle models. Experimental results show that 3DGAA achieves to reduce the detection mAP from 87.21\% to 7.38\%, significantly outperforming existing 3D physical attacks. Moreover, our method maintains high transferability across different physical conditions, demonstrating a new state-of-the-art in physically realizable adversarial attacks.

Feng Zhou, Quyu Kong, Jie Qiao et al.

Temporal point processes (TPPs) are stochastic process models used to characterize event sequences occurring in continuous time. Traditional statistical TPPs have a long-standing history, with numerous models proposed and successfully applied across diverse domains. In recent years, advances in deep learning have spurred the development of neural TPPs, enabling greater flexibility and expressiveness in capturing complex temporal dynamics. The emergence of large language models (LLMs) has further sparked excitement, offering new possibilities for modeling and analyzing event sequences by leveraging their rich contextual understanding. This survey presents a comprehensive review of recent research on TPPs from three perspectives: Bayesian, deep learning, and LLM approaches. We begin with a review of the fundamental concepts of TPPs, followed by an in-depth discussion of model design and parameter estimation techniques in these three frameworks. We also revisit classic application areas of TPPs to highlight their practical relevance. Finally, we outline challenges and promising directions for future research.

Jackie Ayoub, Lilit Avetisyan, Mustapha Makki et al.

Understanding how trust is built over time is essential, as trust plays an important role in the acceptance and adoption of automated vehicles (AVs). This study aimed to investigate the effects of system performance and participants' trust preconditions on dynamic situational trust during takeover transitions. We evaluated the dynamic situational trust of 42 participants using both self-reported and behavioral measures while watching 30 videos with takeover scenarios. The study was a 3 by 2 mixed-subjects design, where the within-subjects variable was the system performance (i.e., accuracy levels of 95\%, 80\%, and 70\%) and the between-subjects variable was the preconditions of the participants' trust (i.e., overtrust and undertrust). Our results showed that participants quickly adjusted their self-reported situational trust (SST) levels which were consistent with different accuracy levels of system performance in both trust preconditions. However, participants' behavioral situational trust (BST) was affected by their trust preconditions across different accuracy levels. For instance, the overtrust precondition significantly increased the agreement fraction compared to the undertrust precondition. The undertrust precondition significantly decreased the switch fraction compared to the overtrust precondition. These results have important implications for designing an in-vehicle trust calibration system for conditional AVs.

Gandhimathi Padmanaban, Nathaniel Jachim, Hala Shandi et al.

Automated driving system - dedicated vehicles (ADS-DVs), specially designed for people with various disabilities, can be beneficial to improve their mobility. However, research related to autonomous vehicles (AVs) for people with cognitive disabilities, especially Autism Spectrum Disorder (ASD) is limited. Thus, in this study, we focused on the challenge that we framed: "How might we design an ADS-DV that benefits people with ASD and their caregivers?". In order to address the design challenge, we followed the human-centered design process. First, we conducted user research with caregivers of people with ASD. Second, we identified their user needs, including safety, monitoring and updates, individual preferences, comfort, trust, and reliability. Third, we generated a large number of ideas with brainstorming and affinity diagrams, based on which we proposed an ADS-DV prototype with a mobile application and an interior design. Fourth, we tested both the low-fidelity and high-fidelity prototypes to fix the possible issues. Our preliminary results showed that such an ASD-DV would potentially improve the mobility of those with ASD without worries.

Sue Bai, Dakota Drake Legge, Ashley Young et al.

In this study, we investigated the effectiveness and user acceptance of three external interaction modalities (i.e., visual, auditory, and visual+auditory) in promoting communications between automated vehicle systems (AVS) and pedestrians at a crosswalk through a large number of combined designs. For this purpose, an online survey was designed and distributed to 68 participants. All participants reported their overall preferences for safety, comfort, trust, ease of understanding, usability, and acceptance towards the systems. Results showed that the visual+auditory interaction modality was the mostly preferred, followed by the visual interaction modality and then the auditory one. We also tested different visual and auditory interaction methods, and found that "Pedestrian silhouette on the front of the vehicle" was the best preferred option while middle-aged participants liked "Chime" much better than young participants though it was overall better preferred than others. Finally, communication between the AVS and pedestrians' phones was not well received due to privacy concerns. These results provided important interface design recommendations in identifying better combination of visual and auditory designs and therefore improving AVS communicating their intention with pedestrians.

Yixuan Zhang, Feng Zhou, Zhidong Li et al.

The label bias and selection bias are acknowledged as two reasons in data that will hinder the fairness of machine-learning outcomes. The label bias occurs when the labeling decision is disturbed by sensitive features, while the selection bias occurs when subjective bias exists during the data sampling. Even worse, models trained on such data can inherit or even intensify the discrimination. Most algorithmic fairness approaches perform an empirical risk minimization with predefined fairness constraints, which tends to trade-off accuracy for fairness. However, such methods would achieve the desired fairness level with the sacrifice of the benefits (receive positive outcomes) for individuals affected by the bias. Therefore, we propose a Bias-TolerantFAirRegularizedLoss (B-FARL), which tries to regain the benefits using data affected by label bias and selection bias. B-FARL takes the biased data as input, calls a model that approximates the one trained with fair but latent data, and thus prevents discrimination without constraints required. In addition, we show the effective components by decomposing B-FARL, and we utilize the meta-learning framework for the B-FARL optimization. The experimental results on real-world datasets show that our method is empirically effective in improving fairness towards the direction of true but latent labels.

Feng Zhou, Areen Alsaid, Mike Blommer et al.

Research indicates that monotonous automated driving increases the incidence of fatigued driving. Although many prediction models based on advanced machine learning techniques were proposed to monitor driver fatigue, especially in manual driving, little is known about how these black-box machine learning models work. In this paper, we proposed a combination of eXtreme Gradient Boosting (XGBoost) and SHAP (SHapley Additive exPlanations) to predict driver fatigue with explanations due to their efficiency and accuracy. First, in order to obtain the ground truth of driver fatigue, PERCLOS (percentage of eyelid closure over the pupil over time) between 0 and 100 was used as the response variable. Second, we built a driver fatigue regression model using both physiological and behavioral measures with XGBoost and it outperformed other selected machine learning models with 3.847 root-mean-squared error (RMSE), 1.768 mean absolute error (MAE) and 0.996 adjusted $R^2$. Third, we employed SHAP to identify the most important predictor variables and uncovered the black-box XGBoost model by showing the main effects of most important predictor variables globally and explaining individual predictions locally. Such an explainable driver fatigue prediction model offered insights into how to intervene in automated driving when necessary, such as during the takeover transition period from automated driving to manual driving.

Na Du, X. Jessie Yang, Feng Zhou

In SAE Level 3 automated driving, taking over control from automation raises significant safety concerns because drivers out of the vehicle control loop have difficulty negotiating takeover transitions. Existing studies on takeover transitions have focused on drivers' behavioral responses to takeover requests (TORs). As a complement, this exploratory study aimed to examine drivers' psychophysiological responses to TORs as a result of varying non-driving-related tasks (NDRTs), traffic density and TOR lead time. A total number of 102 drivers were recruited and each of them experienced 8 takeover events in a high fidelity fixed-base driving simulator. Drivers' gaze behaviors, heart rate (HR) activities, galvanic skin responses (GSRs), and facial expressions were recorded and analyzed during two stages. First, during the automated driving stage, we found that drivers had lower heart rate variability, narrower horizontal gaze dispersion, and shorter eyes-on-road time when they had a high level of cognitive load relative to a low level of cognitive load. Second, during the takeover transition stage, 4s lead time led to inhibited blink numbers and larger maximum and mean GSR phasic activation compared to 7s lead time, whilst heavy traffic density resulted in increased HR acceleration patterns than light traffic density. Our results showed that psychophysiological measures can indicate specific internal states of drivers, including their workload, emotions, attention, and situation awareness in a continuous, non-invasive and real-time manner. The findings provide additional support for the value of using psychophysiological measures in automated driving and for future applications in driver monitoring systems and adaptive alert systems.

Feng Zhou, Tao Chen, Baiying Lei

Amid the ongoing COVID-19 pandemic, whether COVID-19 patients with high risks can be recovered or not depends, to a large extent, on how early they will be treated appropriately before irreversible consequences are caused to the patients by the virus. In this research, we reported an explainable, intuitive, and accurate machine learning model based on logistic regression to predict the fatality rate of COVID-19 patients using only three important blood biomarkers, including lactic dehydrogenase, lymphocyte (%) and high-sensitivity C-reactive protein, and their interactions. We found that when the fatality probability produced by the logistic regression model was over 0.8, the model had the optimal performance in that it was able to predict patient fatalities more than 11.30 days on average with maximally 34.91 days in advance, an accumulative f1-score of 93.76% and and an accumulative accuracy score of 93.92%. Such a model can be used to identify COVID-19 patients with high risks with three blood biomarkers and help the medical systems around the world plan critical medical resources amid this pandemic.

Na Du, Feng Zhou, Elizabeth Pulver et al.

In conditionally automated driving, drivers have difficulty in takeover transitions as they become increasingly decoupled from the operational level of driving. Factors influencing takeover performance, such as takeover lead time and the engagement of non-driving related tasks, have been studied in the past. However, despite the important role emotions play in human-machine interaction and in manual driving, little is known about how emotions influence drivers takeover performance. This study, therefore, examined the effects of emotional valence and arousal on drivers takeover timeliness and quality in conditionally automated driving. We conducted a driving simulation experiment with 32 participants. Movie clips were played for emotion induction. Participants with different levels of emotional valence and arousal were required to take over control from automated driving, and their takeover time and quality were analyzed. Results indicate that positive valence led to better takeover quality in the form of a smaller maximum resulting acceleration and a smaller maximum resulting jerk. However, high arousal did not yield an advantage in takeover time. This study contributes to the literature by demonstrating how emotional valence and arousal affect takeover performance. The benefits of positive emotions carry over from manual driving to conditionally automated driving while the benefits of arousal do not.

Feng Zhou, Shu Kong, Charless Fowlkes et al.

Automated facial expression analysis has a variety of applications in human-computer interaction. Traditional methods mainly analyze prototypical facial expressions of no more than eight discrete emotions as a classification task. However, in practice, spontaneous facial expressions in naturalistic environment can represent not only a wide range of emotions, but also different intensities within an emotion family. In such situation, these methods are not reliable or adequate. In this paper, we propose to train deep convolutional neural networks (CNNs) to analyze facial expressions explainable in a dimensional emotion model. The proposed method accommodates not only a set of basic emotion expressions, but also a full range of other emotions and subtle emotion intensities that we both feel in ourselves and perceive in others in our daily life. Specifically, we first mapped facial expressions into dimensional measures so that we transformed facial expression analysis from a classification problem to a regression one. We then tested our CNN-based methods for facial expression regression and these methods demonstrated promising performance. Moreover, we improved our method by a bilinear pooling which encodes second-order statistics of features. We showed such bilinear-CNN models significantly outperformed their respective baselines.