Using facial recognition as a use-case scenario, we attempt to identify sources of bias in a model developed using transfer learning. To achieve this task, we developed a model based on a pre-trained facial recognition model, and scrutinized the accuracy of the model’s image classification against factors such as age, gender, and race to observe whether or not the model performed better on some demographic groups than others. By identifying the bias and finding potential sources of bias, his work contributes a unique technical perspective from the view of a small scale developer to emerging discussions of accountability and transparency in AI.
Here, recognizing the difficulty associated with tracking the progression of dementia, the authors used machine learning models to predict between the presence of cognitive normalcy, mild cognitive impairment, and Alzheimer's Disease, based on blood DNA methylation levels, sex, and age. With four machine learning models and two dataset dimensionality reduction methods they achieved an accuracy of 53.33%.
The authors trained a machine learning model to detect kidney stones based on characteristics of urine. This method would allow for detection of kidney stones prior to the onset of noticeable symptoms by the patient.
The diagnosis of malaria remains one of the major hurdles to eradicating the disease, especially among poorer populations. Here, the authors use machine learning to improve the accuracy of deep learning algorithms that automate the diagnosis of malaria using images of blood smears from patients, which could make diagnosis easier and faster for many.
There are believed to be ~20,000 nebulae in the Milky Way Galaxy. However, humans have only cataloged ~1,800 of them even though we have gathered 1.3 million nebula images. Classification of nebulae is important as it helps scientists understand the chemical composition of a nebula which in turn helps them understand the material of the original star. Our research on nebulae classification aims to make the process of classifying new nebulae faster and more accurate using a hybrid of deep learning and machine learning techniques.
In this study, the authors test whether providing a larger dataset of glass transition temperatures (Tg) to train the machine-learning platform Polymer Genome would improve its accuracy. Polymer Genome is a machine learning based data-driven informatics platform for polymer property prediction and Tg is one property needed to design new polymers in silico. They found that training the model with their larger, curated dataset improved the algorithm's Tg, providing valuable improvements to this useful platform.
In this paper, we measured the privacy budgets and utilities of different differentially private mechanisms combined with different machine learning models that forecast traffic congestion at future timestamps. We expected the ANNs combined with the Staircase mechanism to perform the best with every value in the privacy budget range, especially with the medium high values of the privacy budget. In this study, we used the Autoregressive Integrated Moving Average (ARIMA) and neural network models to forecast and then added differentially private Laplacian, Gaussian, and Staircase noise to our datasets. We tested two real traffic congestion datasets, experimented with the different models, and examined their utility for different privacy budgets. We found that a favorable combination for this application was neural networks with the Staircase mechanism. Our findings identify the optimal models when dealing with tricky time series forecasting and can be used in non-traffic applications like disease tracking and population growth.
Alzheimer’s disease (AD) is a common disease affecting 6 million people in the U.S., but no cure exists. To create therapy for AD, it is critical to detect amyloid-β protein in the brain at the early stage of AD because the accumulation of amyloid-β over 20 years is believed to cause memory impairment. However, it is difficult to examine amyloid-β in patients’ brains. In this study, we hypothesized that we could accurately predict the presence of amyloid-β using EEG data and machine learning.
With advancements in machine learning a large data scale, high throughput virtual screening has become a more attractive method for screening drug candidates. This study compared the accuracy of molecular descriptors from two cheminformatics Mordred and PaDEL, software libraries, in characterizing the chemo-structural composition of 53 compounds from the non-nucleoside reverse transcriptase inhibitors (NNRTI) class. The classification model built with the filtered set of descriptors from Mordred was superior to the model using PaDEL descriptors. This approach can accelerate the identification of hit compounds and improve the efficiency of the drug discovery pipeline.
Coral bleaching is a fatal process that reduces coral diversity, leads to habitat loss for marine organisms, and is a symptom of climate change. This process occurs when corals expel their symbiotic dinoflagellates, algae that photosynthesize within coral tissue providing corals with glucose. Restoration efforts have attempted to repair damaged reefs; however, there are over 360,000 square miles of coral reefs worldwide, making it challenging to target conservation efforts. Thus, predicting the likelihood of bleaching in a certain region would make it easier to allocate resources for conservation efforts. We developed a machine learning model to predict global locations at risk for coral bleaching. Data obtained from the Biological and Chemical Oceanography Data Management Office consisted of various coral bleaching events and the parameters under which the bleaching occurred. Sea surface temperature, sea surface temperature anomalies, longitude, latitude, and coral depth below the surface were the features found to be most correlated to coral bleaching. Thirty-nine machine learning models were tested to determine which one most accurately used the parameters of interest to predict the percentage of corals that would be bleached. A random forest regressor model with an R-squared value of 0.25 and a root mean squared error value of 7.91 was determined to be the best model for predicting coral bleaching. In the end, the random model had a 96% accuracy in predicting the percentage of corals that would be bleached. This prediction system can make it easier for researchers and conservationists to identify coral bleaching hotspots and properly allocate resources to prevent or mitigate bleaching events.