2731 papers
This paper introduces Causal Adjoint Transport (CAT), a local second-order adjoint dynamics method that significantly reduces the computational cost of training implicit and recurrent neural networks near stability boundaries by requiring substantially fewer Jacobian actions compared to first-order relaxation and other solvers.
This study evaluates current limitations in Medical Visual Question Answering (Med-VQA) for telemedicine and proposes the TRACE-MedVQA framework, a retrieval-augmented, calibrated, and explainable system designed to enhance the safety and reliability of remote clinical decision support.
This paper proposes a retrieval-augmented captioning framework that injects distilled crash-avoidance policies and formal driving rules into Vision-Language-Action models, significantly improving autonomous vehicle trajectory prediction accuracy to match ground-truth performance by addressing safety-critical details often missed by standard models.
This study demonstrates that carefully structured large language model workflows, particularly those employing few-shot prompting and explicit reasoning, can achieve high sensitivity in systematic review screening while reducing manual effort by over 80% at a low cost.
This paper proposes a Condition-Aware Multiscale Hourglass-Transformer (CA-MHT) model that improves Remaining Useful Life prediction for turbofan engines under varying operating conditions through regime-specific normalization and multiscale temporal analysis, while utilizing SHAP and Integrated Gradients to validate the model's reliance on temporal sensor behavior and ensure decision-making transparency.
This paper introduces the "Ocean State," a method that replaces standard self-attention with a single persistent vector to achieve O(1) computation and memory complexity while demonstrating superior training stability and no signs of forgetting even at 10 million steps of context.
This paper presents NumaRing, a topology-aware MPMC queue implementation that demonstrates how profiling-driven discoveries—specifically eliminating a costly per-operation topology lookup, fixing a shared-atomic bottleneck in work-stealing, and removing ineffective CPU-pause backoff—can drastically improve performance, while also revealing that even with these optimizations, raw throughput on a two-socket system remains far below original design targets.
This study proposes an automated hematological classification system using an Enhanced Randomized Forest (ERF) ensemble model trained on 7,196 CBC patient records, which outperforms other machine learning algorithms in diagnostic accuracy and utilizes SHAP and LIME techniques to ensure model interpretability for clinical decision support.
This paper identifies preprocessing bottlenecks, specifically SSD bounding-box limitations, as the primary cause of accuracy loss in lightweight driver drowsiness detection systems and proposes a solution combining CLAHE-adaptive preprocessing, a three-zone confidence protocol, and dynamic quantization to achieve 99.0% accuracy with sub-40ms latency on edge devices.
The paper introduces QTRadX-KOA, a hybrid quantum-classical framework that integrates a DenseNet-121 backbone, a lightweight Transformer, and a novel Residual Hybrid Quantum Head to achieve edge-compatible, resource-efficient Knee Osteoarthritis grading with performance comparable to larger classical models while reducing parameter count by approximately 90%.