Volume 199

Osteoarthritis is a chronic degenerative disease that causes knee pain and movement disorder in patients. There is still no drug to treat the disease, so scientists are working on the drugs. To test new drugs, medical images have played a vital role in medical treatment and diagnosis, and their analysis has become the top priority in clinical treatment. In medical image analysis, accurate segmentation of medical images plays a crucial role in its analysis. In recent years, with the continuous development of artificial intelligence, people have tried to use deep learning and convolutional neural networks to segment images to achieve the effect of reducing the workload of doctors. In this study, we do image segmentation using UNet network from three directions, and we use three 2D segmentation model from different directions to rebuild a 3D segmentation model to have an accurate and reproducible volumetric quantification of articular cartilage in the knee.

Show more

Traditional treatments for ocular motility disorders and eye prosthetics face significant limitations in restoring natural, coordinated eye movements. Conditions such as strabismus or eye paralysis alter the delicate balance between the agonist and antagonist eye muscles, resulting in misalignment, poor focus, and restricted movement. Current interventions, including surgeries or static eye prostheses, often fail to replicate the full range of motion or provide the feedback necessary for dynamic visual function. Therefore, the development of Brain-Computer Interfaces for eye movement rehabilitation was developed with the aim to better the experience of eye motion, as well as to be able to enable precisely coordinated eye control for the amputees. This paper conducted a literature review on interface design, signal image conversions, age-related data differences, and the principles of P300 Speller BCI (a system used to show target characters on a computer screen). The question researched is, “How do spellers and motor imagery function and in what scenarios are their applications suitable”. After examining numerous papers, the conclusion is that BCI technology has the potential to restore eye performance for the unsighted, including clear visual feedback and the ability for precise eye coordination.

Show more

Prompt recovery in large language models (LLMs) is critical for understanding their underlying mechanisms and addressing concerns regarding privacy and copyright. Contemporary LLMs typically provide only inference results, making the process of recovering prompts exceedingly challenging and the accuracy of recovery uncertain. To address this issue, the focus is on extracting information related to prompt recovery from a limited amount of output text and maximizing its utility. LLMs use prompts to generate text, with the prompts often containing background information referred to as secret prompts, which are usually not disclosed to users. However, prompt attacks can be employed to uncover these secret prompts by crafting specific input prompts to exploit the LLMs. This study aims to improve the accuracy of recovering secret prompts by designing a method that combines secret prompts obtained through different models, including the Deliberative Prompt Recovery (DORY) model and the Prompt Attack Extraction System. This combined framework demonstrates superior performance in maintaining high accuracy under lower similarity thresholds. The paper highlights the state-of-the-art capabilities of these two recovery approaches, providing an insightful overview of advancements in prompt recovery. Additionally, it proposes a methodology for integrating an ordinary prompt recovery model with a prompt attack extraction system through a combination algorithm to enhance accuracy in prompt recovery. The study concludes by identifying current challenges and outlining future research and development directions in this domain.

Show more

Bias in machine learning datasets and models can pose significant challenges to achieving fairness in real-world applications. In this paper, we look over two methods aimed at mitigating bias in machine learning datasets: “Identifying and Correcting Label Bias in Machine Learning” and “Debiasing made state-of-the-art Revisiting the Simple Seed-based Weak Supervision for Text Classification”. The first method utilizes a mathematical approach that re-weights training samples, addressing label bias by integrating fairness constraints directly for the optimization process. Such examples include demographic parity and equalized odds; iterative training and adjustments with fairness violation penalties establish a balance between accuracy and fairness. The second method presents seed deletion in weak supervision as a way to minimize bias in text classification tasks. By removing specific seed words from pseudo-labeled texts, and data augmentation via random deletion, the model reduces the overreliance on biased features, which improves robustness and generalization. Overall, we evaluated that these methods can achieve improvements in fairness and accuracy across diverse data sets and domains which include: crime prediction, credit scoring, and text classification. Our paper highlights the potential of combining advanced mathematical techniques with preprocessing to mitigate bias in machine learning.

Show more