A Robust Backward Compatibility Metric for Model Retraining

Label corruption, where training samples are mislabeled due to non-expert annotation or adversarial attacks, significantly degrades model performance. Acquiring large, perfectly labeled datasets is costly, and retraining models from scratch is computationally expensive. To address this, we introduce Scaled Activation Projection (SAP), a novel SVD (Singular Value Decomposition)-based corrective machine unlearning algorithm. SAP mitigates label noise by identifying a small subset of trusted samples using cross-entropy loss and projecting model weights onto a clean activation space estimated using SVD on these trusted samples. This process suppresses the noise introduced in activations due to the mislabeled samples. In our experiments, we demonstrate SAP’s effectiveness on synthetic noise with different settings and real-world label noise. SAP applied to the CIFAR dataset with 25% synthetic corruption show upto 6% generalization improvements. Additionally, SAP can improve the generalization over noise robust training approaches on CIFAR dataset by ∼ 3.2% on average. Further, we observe generalization improvements of 2.31% for a Vision Transformer model trained on naturally corrupted Clothing1M.

Increasing portability of computers will require users in the future to access the network regardless of location. Host migration transparency will be an essential feature of wide area network environments. We proposed the concept of virtual network and the propagating cache method to achieve host migration transparency [12]. We also established the feasibility of Virtual Internet Protocol (VIP), derived from Internet Protocol (IP), as an example of a virtual network protocol. In this paper, we define VIP in detail and describe how VIP achieves host migration transparency in IP networks. We also show two VIP implementation approaches: 1) VIP as a sublayer; and 2) VIP as an IP option. By modifying an operating systems kernel in each case, we implement VIP using both approaches. We then compare both implementations in terms of overhead, backward compatibility, and other metrics. The measured performance indicates that VIP can provide host migration transparency in IP networks with negligible overhead.

Substantial effort has been directed toward demonstrating uses of predictive models in health care. However, implementation of these models into clinical practice may influence patient outcomes, which in turn are captured in electronic health record data. As a result, deployed models may affect the predictive ability of current and future models. To estimate changes in predictive model performance with use through 3 common scenarios: model retraining, sequentially implementing 1 model after another, and intervening in response to a model when 2 are simultaneously implemented. Simulation of model implementation and use in critical care settings at various levels of intervention effectiveness and clinician adherence. Models were either trained or retrained after simulated implementation. Admissions to the intensive care unit (ICU) at Mount Sinai Health System (New York, New York) and Beth Israel Deaconess Medical Center (Boston, Massachusetts). 130 000 critical care admissions across both health systems. Across 3 scenarios, interventions were simulated at varying levels of clinician adherence and effectiveness. Statistical measures of performance, including threshold-independent (area under the curve) and threshold-dependent measures. At fixed 90% sensitivity, in scenario 1 a mortality prediction model lost 9% to 39% specificity after retraining once and in scenario 2 a mortality prediction model lost 8% to 15% specificity when created after the implementation of an acute kidney injury (AKI) prediction model; in scenario 3, models for AKI and mortality prediction implemented simultaneously, each led to reduced effective accuracy of the other by 1% to 28%. In real-world practice, the effectiveness of and adherence to model-based recommendations are rarely known in advance. Only binary classifiers for tabular ICU admissions data were simulated. In simulated ICU settings, a universally effective model-updating approach for maintaining model performance does not seem to exist. Model use may have to be recorded to maintain viability of predictive modeling. National Center for Advancing Translational Sciences.

The purpose of the study: development and experimental validation of a methodology for adaptive cyber threat assessment in the telemetry infrastructure of railway transport using a digital twin for safe machine learning model retraining. Research methods: cyberattack scenario modeling, statistical analysis of time series, machine learning techniques, synthetic data generation within a digital twin environment. Results obtained: an architecture for an adaptive cyber threat detection system is proposed, implementing a closed-loop cycle of monitoring, evaluation, and safe model retraining within an isolated digital twin environment. A synthetic dataset was generated to simulate six types of stealthy and evolving cyberattacks targeting the telemetry of traction power systems and SCADA systems in railway infrastructure. Experimental results demonstrate that a static machine learning model suffers severe performance degradation under concept drift, exhibiting extremely low recall. In contrast, the proposed adaptation mechanism—based on the analysis of false negatives and the targeted addition of representative examples to the training set—significantly improves recall while maintaining high precision. Even the inclusion of a small number of new samples (less than 3 % of the original dataset) yields a substantial increase in F1-score, confirming the approach’s effectiveness under strict regulatory constraints that limit direct intervention in real critical infrastructure operations. Scientific novelty: The proposed approach is distinguished by the integration of a digital twin as a secure environment for generating synthetic attack scenarios and retraining the detection model based on false-negative analysis. This enables effective compensation for concept drift without disrupting the operation of the actual railway infrastructure. Contributions: Kotenko I. V. – development of the problem statement and general approach to adaptive assessment of cyber threats in the telemetry infrastructure of railway transport; Kotenko I. V., Saenko I. B. and Mityakov E. S. – analysis of modern research on cyber threat assessment and development of the architecture of the adaptive cyber threat assessment system; Mityakov E. S. – experimental study of the proposed approach; Kotenko I. V., Saenko I. B. and Mityakov E. S. – analysis of the research results.

PRUNE: A patching based repair framework for certifiable and privacy-robust unlearning of neural networks.

Wearable technologies enable continuous monitoring of various health metrics, such as physical activity, heart rate, sleep, and stress levels. A key challenge with wearable data is obtaining quality labels. Unlike modalities like video where the videos themselves can be effectively used to label objects or events, wearable data do not contain obvious cues about the physical manifestation of the users and usually require rich metadata. As a result, label noise can become an increasingly thorny issue when labeling wearable data. In this paper, we propose a novel solution to address noisy label learning, entitled Few-Shot Human-in-the-Loop Refinement (FHLR). Our method initially learns a seed model using weak labels. Next, it fine-tunes the seed model using a handful of expert corrections. Finally, it achieves better generalizability and robustness by merging the seed and fine-tuned models via weighted parameter averaging. We evaluate our approach on four challenging tasks and datasets, and compare it against eight competitive baselines designed to deal with noisy labels. We show that FHLR achieves significantly better performance when learning from noisy labels and achieves state-of-the-art by a large margin, with up to accuracy improvement under symmetric and asymmetric noise. Notably, we find that FHLR is particularly robust to increased label noise, unlike prior works that suffer from severe performance degradation. Our work not only achieves better generalization in high-stakes health sensing benchmarks but also sheds light on how noise affects commonly-used models.

Modality for scenario analysis and maximum likelihood allocation

Mathematical libraries are typically developed for use with the fixed-width data-paths on processors and target common floating-point formats such as IEEE binary32 and binary64 . To address the increasing energy consumption and throughput requirements of HPC, scientific computing and AI applications, libraries and hardware implementations now provide new floating-point formats, allowing mathematical function evaluations with different performance and accuracy trade-offs. In this article we present a methodology and its associated proof-of-concept tool to evaluate the benefits of custom accuracy of mathematical library calls in HPC and scientific computations. First, our tool collects for each call-site of a mathematical function the input- and output-data profile. Then, using a heuristic exploration algorithm, we estimate the minimal required accuracy by rounding the result to lower precisions. The data profile and accuracy measurement per call-site is used to speculatively select the mathematical function implementation with the most appropriate accuracy for a given scenario. We have tested the methodology with the Intel MKL Vector Math ( VM ) library, leveraging the predefined accuracy levels. We demonstrate the benefits of our approach on two real-world applications: SGP4 , a satellite tracking application, and PATMOS , a Monte Carlo neutron transport code. The robustness of the methodology is estimated by measuring the numerical accuracy of the resulting optimized code, against user-defined criteria. We experiment and discuss generalization across data-sets and finally propose a speculative runtime implementation for PATMOS . The experiment provides an insight into the performance improvements achievable by leveraging the control of per-function call-site accuracy-mode execution of the Intel compiler SVML library. We show benefits from 13 to 55 percent in time reduction for the PATMOS use case.

The multiplier bootstrap is a fast and easy-to-implement alternative to the standard bootstrap; it has been used successfully in many statistical contexts. In this paper, resampling methods based on multipliers are proposed in a general framework where one investigates the stochastic behavior of a random vector $$\mathbf {Y}\in \mathbb {R}^d$$YźRd conditional on a covariate $$X \in \mathbb {R}$$XźR. Specifically, two versions of the multiplier bootstrap adapted to empirical conditional distributions are introduced as alternatives to the conditional bootstrap and their asymptotic validity is formally established. As the method walks hand-in-hand with the functional delta method, theory around the estimation of statistical functionals is developed accordingly; this includes the interval estimation of conditional mean and variance, conditional correlation coefficient, Kendall's dependence measure and copula. Composite inference about univariate and joint conditional distributions is also considered. The sample behavior of the new bootstrap schemes and related estimation methods are investigated via simulations and an illustration on real data is provided.

Evaluation of insulin sensitivity temporal prediction by using quantile regression combined with neural network model.

The uniform description of mineral processes by methods of compound distributions of material features

Approximate computing has is taken into account to enhance the accuracy performance trade off in error-tolerant applications. For several applications, multiplication is the key mathematical process. To style an efficient low power multipliers compressors are used. During this paper we describe the way to encode the faulty rows within the compressors truth table. They are encoded using propagate and generate signals. The proposed multipliers will give outstanding performance in image sharpening process and joint photographic group applications by gaining efficient outputs with high quality and low power. With error management, we will observe the relevancy and usefulness of approximate multipliers.

Approximate computing has been considered to enhance the accuracy performance trade off in error-tolerant applications. For several applications, multiplication is considered as the key mathematical process. To style an efficient low power multipliers, compressors are used. The proposed research work describes the way to encode the faulty rows within the compressors truth table. They are encoded by using the propagate and generate signals. The proposed multipliers will give outstanding performance for image sharpening process and joint photographic group applications by gaining effieient outputs with high quality and low power. With error management, the relevancy and usefulness of approximate multipliers will be observed.

This study examines the role of artificial intelligence in enhancing risk-based decision-making across multiple sectors. The rapid adoption of AI has transformed organizational approaches from reactive to proactive risk management; however, gaps remain in governance, ethical readiness, and real-world implementation. Therefore, this study aims to systematically review the application of AI, identify its benefits and challenges, and analyze the relationship between AI capabilities, risk management processes, and decision outcomes. This research employs a systematic review method guided by the PRISMA framework, analyzing recent literature published between 2023 and 2025 from major academic databases. The findings indicate that AI significantly improves predictive accuracy, operational efficiency, and decision quality through advanced data processing techniques such as machine learning and deep learning. AI applications are widely observed in healthcare, finance, and environmental risk management, demonstrating strong performance in prediction and analysis. However, challenges related to data quality, model transparency, ethical risks, and governance limitations persist. This study concludes that AI should be implemented as a socio-technical system supported by robust governance frameworks and human oversight to ensure responsible, transparent, and sustainable risk-based decision-making.

Listeners use monaural spectral cues to localize sound sources in sagittal planes (along the up-down and front-back directions). How sensorineural hearing loss affects the salience of monaural spectral cues is unclear. To simulate the effects of outer-hair-cell (OHC) dysfunction and the contribution of different auditory-nerve fiber types on localization performance, we incorporated a nonlinear model of the auditory periphery into a model of sagittal-plane sound localization for normal-hearing listeners. The localization model was first evaluated in its ability to predict the effects of spectral cue modifications for normal-hearing listeners. Then, we used it to simulate various degrees of OHC dysfunction applied to different types of auditory-nerve fibers. Predicted localization performance was hardly affected by mild OHC dysfunction but was strongly degraded in conditions involving severe and complete OHC dysfunction. These predictions resemble the usually observed degradation in localization performance induced by sensorineural hearing loss. Predicted localization performance was best when preserving fibers with medium spontaneous rates, which is particularly important in view of noise-induced hearing loss associated with degeneration of this fiber type. On average across listeners, predicted localization performance was strongly related to level discrimination sensitivity of auditory-nerve fibers, indicating an essential role of this coding property for localization accuracy in sagittal planes.