The purpose of the study was to determine whether graph-based machine learning techniques, which have increased prevalence in the last few years, can accurately classify data into one of many clusters, while requiring less labeled training data and parameter tuning as opposed to traditional machine learning algorithms. The results determined that the accuracy of graph-based and traditional classification algorithms depends directly upon the number of features of each dataset, the number of classes in each dataset, and the amount of labeled training data used.
The purpose of this study was to determine the necessity of previous non-algorithmic attacks (Adversarial Training) in light of algorithmic defense methods (Gradient Masking and Defensive Distillation) against FGSM attacks. We found a significant increase in image classification accuracy from defense methods with the non-algorithmic defense method compared to ones without. By analyzing the significance with a McNemar test, we determined that the inclusion of non-algorithmic defense methods is still necessary in light of new algorithmic defense methods.
Given an association between nicotine addiction and gene expression, we hypothesized that expression of genes commonly associated with smoking status would have variable expression between smokers and non-smokers. To test whether gene expression varies between smokers and non-smokers, we analyzed two publicly-available datasets that profiled RNA gene expression from brain (nucleus accumbens) and lung tissue taken from patients identified as smokers or non-smokers. We discovered statistically significant differences in expression of dozens of genes between smokers and non-smokers. To test whether gene expression can be used to predict whether a patient is a smoker or non-smoker, we used gene expression as the training data for a logistic regression or random forest classification model. The random forest classifier trained on lung tissue data showed the most robust results, with area under curve (AUC) values consistently between 0.82 and 0.93. Both models trained on nucleus accumbens data had poorer performance, with AUC values consistently between 0.65 and 0.7 when using random forest. These results suggest gene expression can be used to predict smoking status using traditional machine learning models. Additionally, based on our random forest model, we proposed KCNJ3 and TXLNGY as two candidate markers of smoking status. These findings, coupled with other genes identified in this study, present promising avenues for advancing applications related to the genetic foundation of smoking-related characteristics.
Semantic segmentation - labelling each pixel in an image to a specific class- models require large amounts of manually labeled and collected data to train.
In this experiment, the authors modify the heat equation to account for imperfect insulation during heat transfer and compare it to experimental data to determine which is more accurate.
This study investigates skin tone bias in artificial intelligence models used for dermatological disease classification and evaluates a CycleGAN-based data augmentation approach to improve diagnostic performance on darker skin types. We generated synthetic dark-skinned images to enhance dataset diversity and compared model performance before and after augmentation. The results demonstrate that augmentation with synthetic dermatological images can help reduce disparities in diagnostic performance across skin tones, highlighting a practical strategy for improving fairness in dermatology AI systems.
Algorithmic trading has been increasingly used by Americans. In this work, we tested whether including the opening, closing, and highest prices in three supervised learning models affected their performance. Indeed, we found that including all three prices decreased the error of the prediction significantly.
Here the authors sought to use three machine learning models to predict poverty levels in Cambodia based on available household data. They found teat multilayer perceptron outperformed the other models, with an accuracy of 87 %. They suggest that data-driven approaches such as these could be used more effectively target and alleviate poverty.
