The Research Journal

QViSTA: A Novel Quantum Vision Transformer for Early Multi-Stage Alzheimer’s Diagnosis Using Optimized Variational Quantum Circuits

Authors: Aryan Singhal, Hursh Shah

Published: July 2024

Abstract:

Magnetic resonance imaging (MRI) is widely used by neurologists to detect brain abnormalities such as strokes, tumors, and various forms of dementia, including Alzheimer's disease. However, accurately diagnosing the different stages of Alzheimer's disease remains a challenge, with nearly one in five patients misdiagnosed due to symptom overlap with other conditions. This paper introduces QViSTA, a novel hybrid quantum vision transformer (QViT) model that exploits quantum parallelism to improve early diagnosis and differentiation of Alzheimer's disease stages. By integrating quantum variational circuits (VQCs) with vision transformers (ViTs), QViSTA addresses the data scalability and computational efficiency limitations of classical machine learning models. Using a balanced, multi-class dataset of 40,000 MRI images, QViSTA achieved a validation area under the receiver operating characteristic (AUC) of 87.86% and a test AUC of 86.67%, closely matching the performance of a benchmarked classical ViT while reducing feature space by 3.18%. Early and accurate detection of Alzheimer's disease is critical, as it allows for timely interventions that can significantly improve the quality of life for patients and their caregivers. As more hospitals adopt AI for biomedical imaging, QViSTA's innovative approach could dramatically reduce misdiagnosis rates, improve patient outcomes, and reduce costs.

Pediatric Seizure Detection System: A Data-Driven Approach with Predictive and Generative AI Integration

Author: Tanisha Sood

Published: June 2024

Abstract:

The gravity of seizures in pediatric populations cannot be overstated. From the youngest infants to adolescents on the brink of adulthood, seizures disrupt not only neurological functions but also the fabric of daily life. According to the Centers for Disease Control and Prevention, about 470,000 children in the United States have epilepsy, which is a neurological disorder characterized by recurrent seizures. Drawing from personal encounters with seizures, I recognize the critical need for effective seizure detection mechanisms. This study addresses the challenge of detecting seizures in children, especially in environments where constant supervision may not be feasible. Utilizing CAD, 3D design, and machine learning algorithms, this prototype predicts a seizure onset by analyzing bio data from a simulation of possible data sets carefully overviewed by my pediatrician Sara MachMan. The prototype integrates insights from a literature search, including the CDC publication on seizure detection, highlighting brain activity, fever, and body movements as triggers. Based on these triggers and thorough data analysis with a pediatrician, I was able to discover a breakthrough. Despite encountering a 15% error rate from my simulated data analysis, the prototype effectively alerts parents and medical professionals, demonstrating its potential to reduce the unpredictability of seizures and positively impact lives.

Pediatric Seizure Detection System: A Data-Driven Approach with Predictive and Generative AI Integration

Author: Tanisha Sood

Published: June 2024

Abstract:

The gravity of seizures in pediatric populations cannot be overstated. From the youngest infants to adolescents on the brink of adulthood, seizures disrupt not only neurological functions but also the fabric of daily life. According to the Centers for Disease Control and Prevention, about 470,000 children in the United States have epilepsy, which is a neurological disorder characterized by recurrent seizures. Drawing from personal encounters with seizures, I recognize the critical need for effective seizure detection mechanisms. This study addresses the challenge of detecting seizures in children, especially in environments where constant supervision may not be feasible. Utilizing CAD, 3D design, and machine learning algorithms, this prototype predicts a seizure onset by analyzing bio data from a simulation of possible data sets carefully overviewed by my pediatrician Sara MachMan. The prototype integrates insights from a literature search, including the CDC publication on seizure detection, highlighting brain activity, fever, and body movements as triggers. Based on these triggers and thorough data analysis with a pediatrician, I was able to discover a breakthrough. Despite encountering a 15% error rate from my simulated data analysis, the prototype effectively alerts parents and medical professionals, demonstrating its potential to reduce the unpredictability of seizures and positively impact lives.

Utilizing Chitosan Nanoparticles To Deliver Natural Growth Promoters To Optimize Seed Germination In Glycine Max

Author: Praneel Shah

Published: June 2024

Abstract:

This project aims to create biodegradable nanoparticles that deliver zinc ions to the soybean plant using a slow-release mechanism, allowing for accelerated seed germination without introducing synthetic materials into the environment and limiting overfertilization of agricultural land. Nanoparticles were formed using Ionic Gelation through an alginate pre-gel matrix. Once the nanoparticle solution was formed, the solution was then diluted to concentrations of 0.25, 0.5, and 1 mg/mL for further testing. Nanoparticle supplementation promoted shorter, thicker root growth in Glycine Max versus non-supplemented plants, suggesting improved zinc uptake as traditionally applied zinc fertilizers leach through the soil, requiring plants to develop longer roots to access it. Optimal growth was seen at 0.25 mg/mL, where plants doubled root quantity, increased root density by 74%, and shoot length by 25% under normal conditions. Under drought and nutrient deficient conditions, 0.25 mg/mL plants grew similarly to control conditions, suggesting nanoparticles improve water conservation within the plant and uptake of nutrients present in the soil ecosystem.

Multi-Source Data Fusion for Flood Mapping in Response to Climate Change

Author: Advaith Anand

Published: June 2024

Abstract:

In the context of escalating climate change impacts, precise flood mapping has become crucial for effective disaster response and management. We address this challenge by developing a deep learning model based on the U-Net architecture, specifically tailored for rapid and accurate flood extent mapping using Synthetic Aperture Radar (SAR) data from the Copernicus Sentinel-1 mission published as part of IEEE 2024 GRSS data fusion challenge. Our approach leverages the unique capabilities of U-Net for detailed image segmentation, combined with SAR imagery, digital elevation models, land-use patterns, and historical water presence data to differentiate between water and non-water surfaces accurately.

The model was rigorously trained and validated across a comprehensive dataset representing a variety of geographic conditions and flood events. It demonstrated significant advancements over traditional flood mapping techniques, achieving an overall accuracy of 95% and an F1-score of 0.7. These metrics underscore the model's effectiveness in accurately classifying flooded areas.

Furthermore, the model exhibits excellent scalability and adaptability, offering potential applications beyond the initial flood mapping scope to other environmental and disaster management scenarios. Our findings not only contribute to the improvement of flood response

strategies but also underscore the potential of advanced machine learning techniques in enhancing the accuracy and efficiency of environmental monitoring in the face of global climate challenges. This research represents a significant step forward in the application of deep learning to hydrological extremes, providing a robust tool for emergency response planners and climate scientists alike.

Developing a Novel Machine Learning Based Approach for Enhanced Risk Analysis in Israeli-Palestinian Conflict Zones

Author: Tashvi Bansal

Published: June 2024

Abstract:

The Israeli-Palestinian conflict poses significant safety challenges for individuals in affected regions. However, predictive mechanisms can help individuals navigate war areas, enabling proactive decision-making necessary for tasks like traveling to hospitals or obtaining food supplies. This project leverages novel machine learning techniques to conduct geospatial risk analysis in these conflict zones. I initially modeled fatality frequencies from geographic coordinates using regression. However, drawbacks included poor user-friendliness and less-accurate predictions. Thus, I explored classification models. Though the Random-Forest model initially showed promising accuracy, on-field testing revealed significant data imbalances. I then shifted to K-means clustering to preserve data representative variance, but its dependency on fatality frequency data was impractical. Consequently, I shifted to the KNN algorithm for improved data compatibility, but it necessitated comprehensive data augmentation and preprocessing. To address the imbalance, my initial attempt to normalize data by population failed to capture data variance. Thus, I created a multi-phase bootstrapping algorithm to generate data from highly imbalanced data. After evenly generating data with coordinate-offsets using trigonometry to account for Earth’s geometry, the points lacked spatial awareness. To resolve this, I applied the Haversine formula to dynamically adjust risk classifications based on proximity to higher-risk areas. I then integrated the optimal KNN model (96.57% accuracy) using cross-validation metrics into an application to perform on-field testing. Furthermore, the app has the potential to evolve from prototype to global use, which is why I'm engaging with larger companies for feature-integration into their platforms. Ultimately, this project establishes a new standard for peace technology development for global-conflicts.

Development of an Accurate Biological Aging Clock Using Machine Learning Models and DNA Methylation Patterns in Skeletal Muscle

Author: Rishab Perati

Published: June 2024

Abstract: 

Objective: A leading challenge in aging research is measuring age accurately, as monitoring healthy individuals over decades to assess the effects of interventions for the aging process is time and cost-prohibitive. The goal of this work was to develop an accurate epigenetic skeletal muscle- based aging clock using high-dimensional DNA methylation patterns to calculate chronological age. 

Method: Genome-wide DNA methylation values from skeletal muscle tissue samples from 47 individuals with 200,000 methylation sites (CpGs) per sample were used for the study. Feature selection was used to narrow down CpGs. An Elastic Net regression model with a Leave-One- Sample-Out Cross-Validation (LO-SAMPLE-OCV) was used to evaluate feature relevance and model performance. I then implemented an ensemble approach where predictions from the LO- SAMPLE-OCV process served as inputs to another model using ElasticNet regression. This two- step approach leverages the strengths of ElasticNet for both feature selection and meta-learning, aiming to enhance predictive accuracy while also capturing the complex relationships of high- dimensional data. 

Results & Conclusion: The model outperformed the Voisin et. al. method against the performance metrics of median absolute age differences (6.5 years using the Voisin method versus 1.9 years and 2.31 years using the LO-SAMPLE-OCV and ensemble methods respectively) and mean age differences (1.9 years using the Voisin method versus 0.26 years and -0.17 years using the LO- SAMPLE-OCV and ensemble methods respectively). 

Impact: This tool can be used to accurately predict and understand aging to develop interventions for the prevention, early detection, diagnosis, and treatment of aging-related diseases, including cancer.

A Deep Learning Pipeline for Drought Assessment using Spatial Satellite Images and Vision Transformers

Author: Aaryan Doshi

Published: June 2024

Abstract: 700 million people are in danger of being displaced due to inept drought prediction and prevention systems. Current research on drought assessment focuses solely on factors such as soil moisture and rainfall, which require painstaking measurements and lab samples, and can often be misleading. This research eliminates this requirement by proposing an end-to-end pipeline to detect and prevent droughts in at-risk areas using satellite images and vision transformers. The dataset is comprised of over 86,000 satellite images labeled by pastoralists and divided with an 80-20 ratio for training and validation. First, using feature filtering, normalization, and a Gaussian filter, the images in the dataset are modified to yield a better performance. Next, a deep vision transformer model with multi-headed attention is constructed, consisting of four heads, three transformer layers, and a patch size of five. The final MLP head produces logits for drought severity prediction level. Overall, the best transformer model achieves 78.3% accuracy in predicting drought conditions on a validation set of 10,000, unseen satellite images. In addition, this method outperforms state-of-the-art convolutional neural networks on this classification task, as compared to VGG- 16, ResNet-50 and DenseNet-121 models. The model harnesses AWS cloud computing, deep vision transformers, and specific image augmentation to achieve state-of-the-art results in drought prediction and prevention. With this research, scientists have the potential to assess droughts quickly and accurately, revolutionizing our ability to provide resources and care to those affected by the increasingly common droughts caused by the climate crisis worldwide.

Detecting Music Enjoyment Level Using Electroencephalogram Signals and Machine Learning Techniques

Author: Raymond Feng

Published: June 2024

Abstract: An electroencephalogram (EEG) is a non-invasive technique that records electrical activity in the brain using scalp electrodes. Researchers have studied the use of EEG to detect emotions and moods by collecting signals from participants and analyzing how those signals correlate with their activities. In this study, researchers investigated the relationship between EEG signals and music enjoyment. Participants listened to music while EEG signals were collected. During the signal-processing phase, power spectral densities (PSDs) were computed from the signals, and dominant brainwave frequencies were extracted from the PSDs to form a comprehensive feature matrix. A machine learning approach was then taken to find correlations between the processed data and the music enjoyment level indicated by the participants. To improve on previous research, multiple machine learning models were employed, including K-Nearest Neighbors Classifier, Support Vector Classifier, and Decision Tree Classifier. Hyperparameters were used to fine-tune each model to further increase its performance. The experiments showed that a strong correlation exists, with the Decision Tree Classifier with hyperparameters yielding 85% accuracy. This study proves that EEG is a reliable means to detect music enjoyment and has future applications including personalized music recommendation, mood adjustment, and mental health therapy.

PLAPT: PROTEIN-LIGAND BINDING AFFINITY PREDICTION USING PRE-TRAINED TRANSFORMERS

Authors: Tyler Rose, Nicolò Monti, Navvye Anand, Tianyu Shen

Published: June 2024

Abstract: Predicting protein-ligand binding affinity is crucial for drug discovery, as it enables efficient identification of drug candidates. We introduce PLAPT, a novel model utilizing transfer learning from pre-trained transformers like ProtBERT and ChemBERTa to predict binding affinities with high accuracy. Our method processes one-dimensional protein and ligand sequences, leveraging a branching neural network architecture for feature integration and affinity estimation. We demonstrate PLAPT’s superior performance through validation on multiple datasets, achieving state-of-the-art results while requiring significantly less computational resources for training compared to existing models. Our findings indicate that PLAPT offers a highly effective and accessible approach for accelerating drug discovery efforts.

WaterGate: An Accessible Computational Model of Flooding Patterns

Authors: Navvye Anand, George Cheng, Tyler Rose

Published: June 2024

Abstract: 240 million people are affected by floods each year, reflecting the urgent need for accessible flood prediction and detection. WaterGate is a computational model that uses geographic elevation data and the rational method to predict flooding patterns, generating an interactive 3D model for user accessibility. Computational hydrology applies numerical methods, machine learning algorithms, and computational simulations to understand, predict, and manage water resources --- including floods. Our project employs computational hydrology by analyzing the structure of river tributaries in 2D through polygon clustering, satellite imaging, and various cleaning protocols. We developed respective tributary tree graphs, morphological graphs, and nodes to create a comprehensive tree and 3D model. Afterward, we examine the morphology of flood plains in 3D space, implementing the rational method (Q = CiA) framework with curated relief plots to predict, model, and visualize flooding elevation. Then, we constructed our stream order analysis, waterline delineation, and statistical analysis to validate our data. Lastly, we modeled different river systems and developed further extensions to increase the applicability of WaterGate to communities around the world.

Inferring Hate Speech Trends for Contemporary Tweets Using a Novel Machine Learning Approach from Supervised Learning Algorithms

Author: Aryan Singhal

Published: July 2023

Abstract: Social media platforms such as Twitter have become ubiquitous in our contemporary society, providing a platform for individuals to express their opinions and engage in discussions on a wide range of topics including those that are neutral and controversial. However, the growing popularity of Twitter has also led to an increase in the prevalence of hate speech, which raises concerns about its impact on individuals and society. This research investigates hate speech trends on Twitter by utilizing supervised machine learning classification, specifically Naive Bayes, and employing Natural Language Processing (NLP) features such as on a range of neutral and controversial topics. The study compares the prevalence of hate speech in these topics and tracks such trends from January 2022 to January 2023. The results show that hate speech was nearly 400% more prevalent in controversial topics than in neutral topics over the course of the year. In addition, this research finds that controversial topics are consistently more vulnerable to hate speech throughout the course of the year when compared to neutral topics. To conduct this study, a Multinomial Naive Bayes classification model was trained on a publicly available Twitter dataset that was specifically labeled for semantic hate speech and achieved an accuracy rate of 94.46%. Ultimately, the higher vulnerability of controversial topics should necessitate policymakers to introduce stricter warnings or frequent policy reminders to platform users. Such changes will foster a respectful and inclusive platform for users, preserving their freedom of expression and encouraging constructive discussions.

Correcting Mislabeled Quasars in Extragalactic Catalogs

Author: Arjun Shrivastava

Published: June 2023

Abstract: Quasars, a type of active galactic nuclei (AGNs), are some of the brightest objects in the universe. They allow astronomers to accurately observe distant objects and look farther back in time, offering researchers a better understanding of our universe’s history. However, the X-ray and extragalactic databases that catalog astronomical objects often mislabel quasars as other objects. Therefore, I seek to improve these classifications by identifying quasars in the Deep Fields component of the Canada-France- Hawaii Telescope Legacy Survey (CFHTLS) database and cross-checking them with the extragalactic catalogs. The typical method of classifying an object as a quasar is via manual visual classification: examining unfolded light curves for star-like objects that exhibit irregular short-term and long-term variations in brightness. Unfortunately, this is time-intensive and prone to human error, so I attempted to accelerate the process and boost accuracy with deep neural networks. I first visually classified data of about 4,000 objects for my training set and included four different filters. After preprocessing, I trained a Long Short-Term Memory (LSTM) neural network with different variations of hyperparameters until I achieved an accuracy of 97.5%. When I ran my model through my data, it identified 992 new quasars, 796 of them being actually quasars while the rest were misclassified, yielding an overall binary accuracy of 80% for the entire CHFTLS Deep Field database. Of the identified quasars, 14.8% of which were new quasars and 83.3% were mislabeled in the NASA/IPAC Extragalactic Database (NED). Many of the mislabeled quasars tended to appear as galaxies as well as unidentified sources of ultraviolet or X-ray radiation. In the future, I seek to improve my model for higher accuracy.

Pillar-Based Overhang Generation to Reduce Waste in 3D Printing

Author: Raymond Feng

Published: June 2023

Abstract: Pillar-Based Overhang Generation, or PBOG, is a novel algorithm that generates support structures for 3D printing. In 3D printing, overhangs are parts in midair that cannot be printed. The most popular way to overcome this issue is to add support structure. The PBOG algorithm is proposed to decrease material usage in support structure. The algorithm uses three steps: overhang detection, vertex simplification, and pillar generation. Compared with Cura and Meshmixer, results show that PBOG uses less waste and has a higher success rate. In conclusion, PBOG outperforms currently popular methods Cura and Meshmixer in material waste and success rate.

Deep Learning Pose Estimation Model for Parkisonism and Levodopa-Induced Dyskinesia

Author: Yashnil Saha

Published: June 2023

Abstract: Diagnosing Parkinson’s disease is one of the largest challenges healthcare systems face due to the absence of a specific test for the condition and symptoms varying widely from person to person. Designing an automated model to aid in early diagnosis would greatly contribute to solving this problem. Currently, diagnosis for PD relies on clinical evaluation which has an error rate of approximately 20%, indicating the urgent need for an automated system to be developed. Levodopa is used for the treatment of Parkinson’s Disease (PD) but can lead to motor complications known as levodopa-induced dyskinesia (LID) when taken for too long. PD and LID are evaluated according to the Unified Parkinson’s Disease Rating Scale (UPDRS) and Unified Dyskinesia Rating Scale (UDysRS) scales, respectively, which range from 0 to 4 (0-normal, 4-severely impaired). The tests are conducted by medical personnel and are very subjective. The goal of this project was to design an algorithm using deep learning for assessment of parkinsonism and LID using pose estimation. Two models were created: a regression model to predict the clinical rating from 0 to 4 and a classification model to determine whether the patient had PD or LID. During the feature extraction process, 32 features were extracted per joint trajectory including 15 kinematic, 16 spectral, and the convex hull of the movements. Then, the two neural network models were trained on these features to be able to predict their respective targets. The classification model achieved a mean F1-score greater than 0.8 and the regression model attained a root mean square error less than 0.550, proving that this project was a promising start in the venture to automate diagnosis of Parkinson’s disease.

Detection of Parkinson's disease using Breathing Signals

Author: Advaith Anand

Published: June 2023

Abstract: Parkinson's disease (PD) is a progressive neurodegenerative disorder that affects millions of people worldwide. The disease primarily impacts the dopaminergic neurons in the substantia nigra, leading to motor symptoms such as tremors, muscle rigidity, and loss of balance. Early diagnosis of PD is crucial for better management of the disease, allowing for early initiation of treatment and improved patient outcomes. Currently, there is no definitive test for PD; diagnosis relies primarily on the evaluation of clinical symptoms, which often appear several years after the onset of the disease. This delayed detection limits the potential for early intervention and disease management. As a result, there is a growing need for non-invasive, cost-effective, and reliable methods to detect PD at an early stage. In this study, I propose a novel approach for early PD detection by capturing respiratory breathing patterns during sleep using smartphone-generated ultrasonic rays. I employ a Deep Neural Network (DNN) model combining a convolutional neural network (CNN) and a recurrent neural network (RNN) for the classification of the breathing signals into two classes: PD and control.

Deep-Learning Based Automatic Ergonomic Assessment Using Webcam Data

Author: Owen Lu

Published: June 2023

Abstract: Primarily due to increasing computer use, people are spending more and more time sitting in front of a desk every day. However, prolonged sitting has been associated with tiredness, hypertension, and pain in areas like the lower back or shoulders. These symptoms may arise for a variety of reasons, but musculoskeletal disorders in particular are largely associated with poor posture. Nevertheless, the adverse results caused by poor posture can be controlled with proper training and monitoring. Thus, this study attempts to provide automatic ergonomic assessment using only image data. Since laptops, desktops, and phones are now widely available and equipped with built-in cameras, this solution is accessible and convenient for most people. More importantly, automatic posture assessment may help prevent conditions associated with poor posture by giving reminders whenever improper posture occurs. To create our model, we make use of Mediapipe, which provides a solution to identifying keypoint locations from an image. By training a fully connected neural network (FC) on this keypoint data, we achieved a 96.96% test F1 score, indicating that our system serves as a convenient way to assess posture while maintaining high performance. To illustrate our results, we perform a final video classification by overlaying the model’s predictions on each frame.

Particle Geodesics in the Kerr Spacetime

Author: Jenna Van Dyke

Published: June 2022

Abstract: The Kerr spacetime is investigated. The singularity structure and horizon structure of the spacetime are explored. Two sets of equations of motion for a particle orbiting a Kerr black hole are derived: the first by the calculating nonzero Christoffel symbols of the metric and utilizing the geodesic equation; the second by dot products of Killing vectors and the four-velocity. The effective potential of the radial motion is found and analyzed in comparison with the Schwarzschild effective potential. The Kerr/CFT correspondence and its recent use in predicting polarimetric images of the supermassive black hole in the galaxy M87 is also discussed briefly. An overview of special relativity, general relativity, and orbits is provided in the appendix to serve as background information for the body of the paper. The Mathematica programs used for this paper can be found here: https://bit.ly/3dlwYH7.

Leveraging Machine Learning and Model Agnostic Explanations to Understand Automated Diagnosis of Cardiovascular Disease

Authors: Christopher Sun, Jai Sharma, Milind Maiti

Published: June 2022

Abstract: The pervasiveness of cardiovascular disease and physician misdiagnosis creates the urgent need for artificial intelligence models to improve diagnosis accuracy. The first objective of this study was to train machine learning models on publicly available data sets containing simple medical information of patients to diagnose cardiovascular disease. The Multilayer Perceptron (MLP) assembled for this task performed optimally with an F1 score of 0.8968. This prompted the creation of an open-source, automated cardiovascular disease diagnosis tool, powered by the MLP. The second objective of this study was to employ a meta-learning methodology called Local Interpretable Model-Agnostic Explanations (LIME) to understand the impact of different features on the model’s diagnosis in the form of marginal probabilities. K-Means Clustering was employed to segment the data into ten clusters, after which each data example was passed through LIME. The resulting histograms depict the complex relationship between feature, cluster, and impact on diagnosis. A series of P-values with contrasting orders of magnitude shows the nuances in the MLP’s understanding of patients from different clusters. The results of meta-learning analysis reveal that the most important features for cardiovascular disease diagnosis are fasting blood sugar, type of chest pain, and slope of the ST segment on an electrocardiogram. Future experiments should replicate the novel methodology introduced in this study on data sets containing more specialized medical features in order to gain practical medical insights about different types of cardiovascular disease represented by each cluster. Finally, feature engineering pathways should be explored with consideration of these results to create versatile diagnosis models not only for cardiovascular disease, but adaptable to other diseases as well.

A machine learning approach to exoplanet prediction through pattern analysis using the transit method

Author: Amy Zheng

Published: June 2022

Analyzing Multispectral Satellite Imagery of South American Wildfires Using Deep Learning

Author: Christopher Sun

Published: June 2022

Abstract: Since frequent severe droughts are lengthening the dry season in the Amazon Rainforest, it is important to detect wildfires promptly and forecast possible spread for effective suppression response. Current wildfire detection models are not versatile enough for the low-technology conditions of South American hot spots. This deep learning study first trains a Fully Convolutional Neural Network on Landsat 8 images of Ecuador and the Galapagos, using Green and Short-wave Infrared bands to predict pixel-level binary fire masks. This model achieves a 0.962 validation F2 score and a 0.932 F2 score on test data from Guyana and Suriname. Afterward, image segmentation is conducted on the Cirrus band using K-Means Clustering to simplify continuous pixel values into three discrete classes representing differing degrees of cirrus cloud contamination. Three additional Convolutional Neural Networks are trained to conduct a sensitivity analysis measuring the effect of simplified features on model accuracy and train time. The Experimental model trained on the segmented cirrus images provides a statistically significant decrease in train time compared to the Control model trained on raw cirrus images, without compromising binary accuracy. This proof of concept reveals that feature engineering can improve the performance of wildfire detection models by lowering computational expense.

A Deep Learning Ensemble Framework for Off-Nadir Geocentric Pose Prediction

Authors: Christopher Sun, Jai Sharma, Milind Maiti

Published: June 2022

Abstract: Roughly 6,800 natural disasters occur annually, and this alarming number continues to grow due to climate change. Effective methods to improve natural disaster response include change detection, map alignment, and vision-aided navigation to allow for the time-efficient delivery of life-saving aid. Current software functions optimally only on near-nadir images taken around ninety degrees above ground level. The inability to generalize to oblique images increases the need to compute an image’s geocentric pose, which is its spatial orientation with respect to gravity. This Deep Learning investigation presents three convolutional models to predict geocentric pose using 5,923 nadir and off-nadir red, green, and blue (RGB) satellite images of cities worldwide. Prior to the ensemble, an autoencoder is assembled to condense the 256 × 256 × 3 images to 32 × 32 × 16 latent space representations, demonstrating the ability to learn useful features from the data. The first model in the ensemble is a U-Net Fully Convolutional Network (FCN) with skip connections used to predict each image’s corresponding pixel-level above-ground elevation mask. This FCN achieves a median absolute deviation of 0.335 meters and an R2 of 0.865 on test data. Afterward, the elevation masks are concatenated with the RGB images to form four-channel inputs fed into the second model in the ensemble, which predicts each image’s rotation angle and magnification scale, the components of its geocentric pose. This Deep Convolutional Model achieves an R2 of 0.943 on test data, significantly outperforming previous models designed by researchers. In addition to achieving superior model performance, outlier removal was performed through supervised interpolation, and the usefulness of data features was gauged through a sensitivity analysis of elevation masks, to target future avenues of feature engineering. The high-accuracy software built in this study contributes to mapping and navigation procedures to accelerate disaster relief and save human lives.

The Effect of Iron as a Potential Inducer of Cataracts

Author: Ethan Liu

Published: June 2022

Abstract: Cataracts are a leading cause of blindness, especially in the developing world. With clean water not always accessible in many third-world countries, the poor quality of water may trigger oxidative stress, a major inducer of cataracts, as demonstrated by previous studies, and one-way oxidation can occur is through Fenton reactions reacting with iron. Perhaps the iron in the untreated water causes oxidation, and this could be a potential cause of cataracts. The objective of the study is to find if the presence of iron ions in drinkable water sources is a contributing factor to cataracts. Pig eyes were soaked in iron solutions and stored under a UV light for 24 hours, simulating human eyes in contact with iron water over a long period, building up oxidative damage. The lenses were then analyzed by comparing the color of the lenses and the change in luminance compared to the control. The results were found to show how iron in water can induce oxidation in the eye due to the Fenton Reaction, leading to cataracts.

Applying adversarial networks to increase the data efficiency and reliability of Self-Driving Cars

Author: Aakash Kumar

Published: June 2022

Abstract: Convolutional Neural Networks (CNNs) are vulnerable to misclassifying images when small perturbations are present. With the increasing prevalence of CNNs in self-driving cars, it is vital to ensure these algorithms are robust to prevent collisions from occurring due to failure in recognizing a situation. In the Adversarial Self-Driving framework, a Generative Adversarial Network (GAN) is implemented to generate realistic perturbations in an image that cause a classifier CNN to misclassify data. This perturbed data is then used to train the classifier CNN further. The Adversarial Self-driving framework is applied to an image classification algorithm to improve the classification accuracy on perturbed images and is later applied to train a self-driving car to drive in a simulation. A small-scale self-driving car is also built to drive around a track and classify signs. The Adversarial Self-driving framework produces perturbed images through learning a dataset, as a result removing the need to train on significant amounts of data. Experiments demonstrate that the Adversarial Self-driving framework identifies situations where CNNs are vulnerable to perturbations and generates new examples of these situations for the CNN to train on. The additional data generated by the Adversarial Self-driving framework provides sufficient data for the CNN to generalize to the environment. Therefore, it is a viable tool to increase the resilience of CNNs to perturbations. Particularly, in the real-world self-driving car, the application of the Adversarial Self-Driving framework resulted in an 18% increase in accuracy, and the simulated self-driving model had no collisions in 30 minutes of driving. 

Investigating the Relationship Between Dropout Regularization and Model Complexity in Neural Networks

Authors: Jai Sharma, Milind Maiti, Christopher Sun

Published: August 2021

Abstract: Dropout Regularization, serving to reduce variance, is nearly ubiquitous in Deep Learning models. We explore the relationship between the dropout rate and model complexity by training 2,000 neural networks configured with random combinations of the dropout rate and the number of hidden units in each dense layer, on each of the three data sets we selected. The generated figures, with binary cross entropy loss and binary accuracy on the z-axis, question the common assumption that adding depth to a dense layer while increasing the dropout rate will certainly enhance performance. We also discover a complex correlation between the two hyperparameters that we proceed to quantify by building additional machine learning and Deep Learning models which predict the optimal dropout rate given some hidden units in each dense layer. Linear regression and polynomial logistic regression require the use of arbitrary thresholds to select the cost data points included in the regression and to assign the cost data points a binary classification, respectively. These machine learning models have mediocre performance because their naive nature prevented the modeling of complex decision boundaries. Turning to Deep Learning models, we build neural networks that predict the optimal dropout rate given the number of hidden units in each dense layer, the desired cost, and the desired accuracy of the model. Though, this attempt encounters a mathematical error that can be attributed to the failure of the vertical line test. The ultimate Deep Learning model is a neural network whose decision boundary represents the 2,000 previously generated data points. This final model leads us to devise a promising method for tuning hyperparameters to minimize computational expense yet maximize performance. The strategy can be applied to any model hyperparameters, with the prospect of more efficient tuning in industrial models.

Identifying Lead-free Perovskites Using Machine Learning for High Efficiency Solar Cells

Author: Sanjana S. Jilla

Published: June 2021

Abstract: One of the biggest threats our planet faces is the threat of global warming. Electrical power has become crucial to our modern lifestyle, but today, electricity is generated by burning fossil fuels and coal, which has many harms and disadvantages associated with it. Fortunately, there are multiple alternative energy sources, including sunlight, a powerful, inexhaustible, and clean resource. There are several different types of solar cells, but one of the most efficient and low-cost types is the perovskite solar cell. Perovskite solar cells (PSCs) have recently received considerable attention due to the high energy conversion efficiency achieved within a few years of their inception. However, today, the most common perovskite is methyl ammonium iodide (MAPbI3), which contains levels of toxic lead. The science community has been searching for lower-toxicity perovskite-type materials, but testing all of the possible lead-free perovskites requires a huge amount of time and funding. Recent advances in computing power have enabled the generation of large datasets for materials and data-driven approaches to problem-solving in materials science, including materials discovery. Machine learning is the primary tool for manipulating such large datasets, predicting unknown material properties and uncovering relationships between structure and property. The goal of this project is to create a Machine Learning (ML) driven software system that increases the efficiency of the solar cell design process. I will do this by identifying the best perovskites by optimizing material composition and determining the importance of the features of each element in the perovskite to the overall efficiency of the PSC. The Machine Learning (ML) driven software system needs to accurately predict key information such as the heat of formation (delta Hf) and band gap (Eg) and accurately use the training data to form an accurate prediction of the best materials to form a perovskite solar cell (Im). The models must predict with 90% accuracy of prediction for the project to be successful. This program is written in Python, and uses Machine Learning to make predictions about the heat of formation and bandgap of various double halide perovskites.

Searching for Rare Quasar Ca II Absorbers in the Early Universe

Author: Iona Xia

Published: June 2021

Abstract: The universe is still largely a mystery to scientists. Quasars are one such mystery, whose emission spectra produce absorption lines such as Ca II when passing through gas of distant galaxies. This data helps astronomers understand more about interstellar gas, dust, and galaxy and star formation and evolution (including my Milky Way). However, these current absorber databases are extremely limited, and traditional methods make them hard to detect. Thus, I seek to discover more Ca II absorbers by developing deep neural networks, which are more accurate and faster. I first found absorbers traditionally to produce a test set. I cropped, normalized, and handpicked through thousands of spectra and discovered 256 Ca II original absorbers. To obtain large training sets, I generated tens of thousands of artificial samples by inserting Ca II lines at corresponding wavelengths in real spectra. I preprocessed the data and created neural network models after testing different hyperparameter configurations. Overall, my accuracy for absorber detection is 95% (Ca II), 15 times higher than traditional methods, and I added significant amounts of new absorbers to the current dataset for Ca II, completing my goal. As for challenges, I concluded that most false negatives are due to noise and weak lines. Furthermore, my discovered absorbers agree with statistical tests of previous studies. In the future, I plan to discover more absorbers using my models and run statistical studies on them.

The Effect of Different Amounts of Potassium Iodide on Asthma in Drosophila Melanogaster

Author: Anika Nagavara

Published: May 2021

Abstract: According to the Centers for Disease Control and Prevention (CDC), 1 in 13 people have asthma. Each day, ten Americans die from asthma (Asthma and Allergy Foundation of America 2018). A previous project that was conducted last year looked at the correlation between zinc deficiency and asthma and yielded positive results meaning that a correlation between additional zinc intake and a better control of asthma could be seen. This project observed whether there is a correlation between iodine deficiency and asthma since iodine has previously been used to treat inflammatory diseases since it has properties that can stabilize thyroid hormone levels as well as reduce bronchial secretions and mediate immune cell responses (Lake 2017). In order to supply the iodine, potassium iodide was used since iodine is more easily absorbed by the body when it is in the form of potassium iodide. A combination treatment of potassium iodide and zinc was also given.

Observing the Genetic Difference of alternative spliced multiple myeloma cells with dose-escalation Tunicamycin

Author: Rishi Pankhaniya

Published: March 2021

Abstract: Tunicamycin is a commonly used drug to cause an unfolded protein response in multiple myeloma cells in order to treat the cancer. However, many multiple myeloma cell lines have slowly developed resistance to this treatment. The goal of my project is to find out the reasons in the RNA behind why this resistance is caused, in order to create a better, altered treatment that could possibly circumvent these problems. First, multiple myeloma cells were treated with tunicamycin repeatedly four times, such that the living cells would sufficiently have developed resistance to the treatment. Then a short-term treatment was performed before plating the cells, dividing the cells into a control and treated group to find differences between their RNA to find indicators that cause the resistance. After plating the cells and isolating the RNA, a mass transcriptome analysis returned exonic data to allow us to look at how the resistance was being developed through the creation of proteins and certain protein responses. Upon looking at the data, several markers were made clear, such as the suppression of VAPA and DDIT3 as examples. Through looking at all of these gene markers, the treatment can be slightly altered in order to prevent the suppression of certain responses that would cause the cancer cell to die, thus

making the treatment apply to a wider range of cells.

Looping and Divergence in the Collatz Conjecture

Authors: Jai Sharma, Akshat Jha, Sambhabi Bose, and Garrett Heller

Published: October 2020