Accelerating Neural Policy Repair with Preservation via Stability-Plasticity Interpolation

2025-05-06

Neural network (NN) control has been adopted widely in cyberphysical systems (CPS). When an NN-based policy fails a formally specified task, NN repair algorithms can fix it. Recent literature raises the problem of Repair with Preservation (RwP), which requires preserving existing correct behaviors while repairing the incorrect ones; a corresponding solution is given, known as Incremental Simulated Annealing Repair (ISAR). In this paper, we tackle the computational efficiency issue of ISAR, which involves expensive log-barriered objective functions and wastes computational efforts rolling back when a repaired NN breaks correct behaviors. With our analysis, we reduce the RwP problem to a stability-plasticity (S-P) trade-off interpolation problem, which has been studied in continual learning (CL). Then, we propose our method, ISAR with Interpolation (ISAR-I), which majorly improves ISAR. ISAR-I abandons the expensive log barriers and rolls back to allow intermediate policies to compromise correct behaviors for repair. Then, an interpolation of the S-P trade-off between the original NN and the intermediate NN is kicked off in the Bayesian space, searching for a final NN that both repairs and preserves. Case studies in OpenAI Gym mountain car and an unmanned underwater vehicle show that ISAR-I is able to preserve all verified trajectories while repairing 81.7% and 21.3% of the broken ones, respectively, achieving the same performance as ISAR, with runtime cost of only 6.5% and 19.6%, on average. Source code: https://github.com/ericlupy/isar_interpolation

VIBRANT: Early Prediction of Life-Threatening Uterine Atony Using Maternal Heart Rate

2025-06-24 · 1 citations

Uterine atony accounts for a vast majority of all postpartum hemorrhages (PPH), the leading cause of maternal mortality worldwide. Uterine atony occurs when the uterine muscle (called the myometrium) does not sufficiently contract to arrest parturient bleeding after delivery. While there exist treatments for uterine atony, delays in intervention reduce their effectiveness. To improve time-to-intervention, postpartum hemorrhage risk prediction tools have been integrated into the standard-of-care for obstetrics. Unfortunately, these tools miss almost half of all PPHs. This paper presents VIBRANT as a clinical decision support tool providing early prediction of potentially life-threatening uterine atony. VIBRANT is physiologically inspired and designed to identify signals of myometrial fatigue argued to be present in streaming maternal heart rate data. Evaluations of the system indicate that VIBRANT, at a clinically actionable specificity, identifies over 80% of potentially life-threatening uterine atony missed by current gold-standard risk prediction tools. Moreover, the system provides between 2 and 8 hours advance warning to care teams prior to delivery for parturients at high-risk, providing ample time for preparation and timely intervention. VIBRANT has been licensed to a commercial partner and is in preparation for a clinical trial to support a regulatory submission and pre-market approval.

Facile Synthesis of a Composite Comprising Multiwalled Carbon Nanotubes with a <i>p</i>–<i>n</i> Heterojunction of Zinc Oxide and Silver Oxide: Applications to Efficient Photocatalytic Decomposition of Emerging Hazardous Pollutants Under Sunlight

ACS Applied Engineering Materials · 2025-06-17 · 1 citations

Bisphenol A (BPA) is a ubiquitous chemical in our daily lives, and its production and use are rapidly increasing, with a high risk of exposure to the environment and living organisms through water pollution. Therefore, we have developed efficient semiconductor-based photocatalysts coupled with multiwalled carbon nanotubes (MWCNTs) to treat and detoxify BPA-containing wastewater in the solar light spectrum. Specifically, to reduce the problems of a binary composite (ZAO) of zinc oxide nanoparticles and silver oxide nanoparticles, MWCNTs were integrated through a wet chemical, cocrystallization, and sealed-state high-pressure annealing process. Morphological investigations revealed that MWCNTs successfully formed a composite with ZAO, a p–n heterojunction, creating a ternary composite (ZnO–Ag2O@MWCNTs, ZAMC). Results showed that exciton recombination was substantially reduced. During 50 min of simulated sunlight exposure, ZAMC demonstrated superior photocatalytic performance, achieving 92.4% photodegradation of BPA. Similarly, the composite further demonstrated robust photocatalytic performance, resulting 96.1%, 92.0%, and 90.9% degradation of a cation dye, methylene blue, and pharmaceuticals, tetracycline and ibuprofen. A recyclability test showed only a 7.7% decrease in performance after five cycles. Similarly, ZAMC showed maximal photocatalytic activity at pH 12. The overall improved performance of ZAMC is credited to a well-developed heterojunction formation that leads to a reduced bandgap, a greater surface area, a greater abundance of active sites, and prolonged retention of the photogenerated excitons.

Tracking Blink Dynamics and Mental States on Glasses

2025-06-23

Eye blink dynamics offer crucial insights into physiological and cognitive states. Existing solutions either require specialized equipment for detailed measurements or sacrifice temporal resolution for accessibility. We present BlinkWise, the first system that transforms everyday eyewear into a wise device for detailed blink dynamics tracking as a tiny add-on. BlinkWise measures eye openness in real-time on resource-constrained edge devices by leveraging both RF modality's inherent efficiency and novel computational optimizations—including recurrentization of convolutional network operations, quantization-aware normalization, and a lightweight proposal algorithm. Evaluation with 20 subjects and more than 18,000 blink measurements demonstrated BlinkWise's high accuracy in capturing subtle blink dynamics at millisecond resolution, achieving a Pearson correlation of 0.981 with ground truth. Through three real-world application studies, we demonstrate BlinkWise's potential for monitoring cognitive states and ocular health. Code, datasets, and demo videos are available on our website.

Ophtimus-V2-Tx: A Compact Domain-Specific LLM for Ophthalmic Diagnosis and Treatment Planning

Research Square · 2025-08-04

Ophtimus-V2-Tx: a compact domain-specific LLM for ophthalmic diagnosis and treatment planning

Scientific Reports · 2025-12-10

Large language models (LLMs) show promise for clinical decision support but often struggle with case-specific reasoning. We present Ophtimus-V2-Tx, an 8-billion-parameter ophthalmology-specialized LLM fine-tuned on more than 10,000 case reports. Evaluation is conducted on a pre-collected dataset. Alongside text metrics (ROUGE-L, BLEU, METEOR) and a semantic similarity score, we use CliBench to map outputs to standardized codes (ICD-10-CM, ATC, ICD-10-PCS) and compute hierarchical F1 (L1-L4 and Full), with code mapping used strictly as an evaluation tool. Ophtimus-V2-Tx is competitive with a state-of-the-art general model and stronger in several settings. It improves text metrics (ROUGE-L 0.40 vs. 0.18; BLEU 0.26 vs. 0.05; METEOR 0.45 vs. 0.29) with comparable semantic similarity. On CliBench, it attains a higher full-code score for secondary diagnosis and ties or leads at selected granular levels for primary diagnosis, while medication and procedure results are close with overlapping confidence intervals. Relative to other ophthalmology-tuned baselines, it shows consistently higher text-generation scores. These findings indicate that a compact, domain-adapted model can approach-or in targeted settings, exceed-large general LLMs on clinically grounded outputs while remaining feasible for on-premise use. We also describe an auditable evaluation pipeline (frozen coding agent, identical prompts, hierarchical metrics) to support reproducibility and future benchmarking.

Conservative Perception Models for Probabilistic Verification

ArXiv.org · 2025-03-23

Verifying the behaviors of autonomous systems with learned perception components is a challenging problem due to the complexity of the perception and the uncertainty of operating environments. Probabilistic model checking is a powerful tool for providing guarantees on stochastic models of systems. However, constructing model-checkable models of black-box perception components for system-level mathematical guarantees has been an enduring challenge. In this paper, we propose a method for constructing provably conservative Interval Markov Decision Process (IMDP) models of closed-loop systems with perception components. We prove that our technique results in conservative abstractions with a user-specified probability. We evaluate our approach in an automatic braking case study using both a synthetic perception component and the object detector YOLO11 in the CARLA driving simulator.

Older Adults’ Perceptions and Attitudes Toward Passive Sensors to Measure Loneliness: A Qualitative Study

Sage Open Aging · 2025-04-29 · 3 citations

Background: Loneliness among older adults significantly affects their health, increasing the risk of mortality, cardiovascular disease, and cognitive decline. Approximately 43% of American adults aged 60 and older report loneliness. Technology-based solutions such as passive sensors offer innovative ways to monitor and alleviate loneliness. In addition to understanding older adults' overall attitudes toward passive sensing, it is important to examine how they may perceive such technologies specifically for the assessment of loneliness. This study examines older adults' perceptions and attitudes toward the use of passive sensors to measure loneliness, with the goal of understanding the feasibility and acceptability of the technology. Methods: A single-group, longitudinal, observational study was conducted with 17 older adults aged 65 and older living independently in Philadelphia. Seven types of sensors were installed in participants' homes for 6 months to monitor motion, proximity, temperature, television viewing, sleep quality, and physical activity. Participants completed the UCLA Loneliness Scale and were interviewed after the study about their experiences with the sensors. Thematic analysis was used to identify key attitudes and perceptions, and SPSS was used for demographic analysis. Results: The mean age of the participants was 73.5 years (range 67-84). Key themes that emerged from the analysis include participants' adaptation and adjustment, privacy and trust in data sharing, and perceived benefits and concerns of the installed sensors. Benefits included detecting and reducing loneliness, monitoring physiological parameters, improving healthier behaviors, etc. Concerns about privacy were raised, such as potential misuse by authorities or third parties, lack of accuracy of functions or feasibility of sensors, and lack of human response. Discussion: Passive sensors were generally accepted, but raised concerns about data security and skepticism about measuring emotional states. Transparent communication and education are needed to address these issues. Improved sensor designs that are more user-centered could encourage adoption, and the technology shows promise for both measuring and potentially alleviating loneliness (e.g., through prompting increased activity).

RICL: Adding In-Context Adaptability to Pre-Trained Vision-Language-Action Models

ArXiv.org · 2025-08-04

Multi-task ``vision-language-action'' (VLA) models have recently demonstrated increasing promise as generalist foundation models for robotics, achieving non-trivial performance out of the box on new tasks in new environments. However, for such models to be truly useful, an end user must have easy means to teach them to improve. For language and vision models, the emergent ability to perform in-context learning (ICL) has proven to be a versatile and highly useful interface to easily teach new tasks with no parameter finetuning. Unfortunately, VLAs pre-trained with imitation learning objectives do not naturally acquire ICL abilities. In this paper, we demonstrate that, with the right finetuning recipe and a small robot demonstration dataset, it is possible to inject in-context adaptability post hoc into such a VLA. After retraining for in-context learning (RICL), our system permits an end user to provide a small number (10-20) of demonstrations for a new task. RICL then fetches the most relevant portions of those demonstrations into the VLA context to exploit ICL, performing the new task and boosting task performance. We apply RICL to inject ICL into the $π_{0}$-FAST VLA, and show that it permits large in-context improvements for a variety of new manipulation tasks with only 20 demonstrations per task, without any parameter updates. When parameter updates on the target task demonstrations is possible, RICL finetuning further boosts performance. We release code and model weights for RICL-$π_{0}$-FAST alongside the paper to enable, for the first time, a simple in-context learning interface for new manipulation tasks. Website: https://ricl-vla.github.io.

Understanding Human Communication Patterns with LLMs: A Study of Short Conversations in Autistic Children

Research Square · 2025-01-28