Undoubtedly, the cornerstone of the project is the exploration of possible improvements for convolutional neural networks. CNNs have been used for object recognition a million times, and having a specific genus of fungi as a subject is nothing special enough to warrant a research paper. However, they do present an interesting issue: they are similar enough to cause a standard CNN issue with identification. Worse still, there is an even bigger host of issues arising from sourcing data from open datasets, such as obscured images, bad positioning, incorrect labeling, and countless others. In short, my project was not only tackling how to create a CNN to classify mushrooms

Ailanthus Altissima, also known as the Tree of Heaven, is an invasive species that causes ecological damage by secreting toxic chemicals into the soil and competing with native species. Aside from ecological damages, it also serves as a host plant for numerous pests and insects, including the spotted lanternfly. The challenge posed by the Tree of Heaven necessitates efforts in management. Through this research study, the Tree of Heaven is monitored remotely using satellite images and machine learning. Due to the absence of studies, surveys, and data on the Tree of Heaven in the United States, a survey dataset conducted in Italy was used to train machine learning models along with satellite imagery to predict the spread of the Tree of Heaven in the United States.

In today’s world, a robotic surgeon manually manipulates the robot via a console using joystick-like controllers. However, such manual methods of control require the surgeon to spend a significant amount of time performing routine but important tasks, such as suturing or tissue cutting, thousands of times over the course of a single surgery. By automating these tasks, we can improve the accuracy and efficiency of the surgery while also allowing the surgeon more time to focus on the more complex surgical tasks at hand.

Growing up reading Isaac Asimov’s robot series, I always envisioned a future where mobile robots will run around helping humans. Of course, this has not happened, and I realized that there are many difficulties to achieving my dream future-cost, energy use, technical difficulties, etc. Although I did not have any significant experience with robotics, I have always been involved with maker technology, both through reading research and news articles and through making smaller-scale gadgets. So, in my project, I wanted to create a robot that addresses some of the most prominent challenges in robotics.

Cardiovascular disease (CVD) is responsible for an estimated 17.9 million annual deaths. Myocardial infarction (MI), a prominent symptom of CVD, occurs when reduced blood flow causes heart muscle death. Since permanent damage to the heart muscle begins within 30 minutes of blood flow restriction, MI is extremely dangerous and time sensitive. Electrocardiogram (ECG) is one of the most efficient methods for MI detection; however, it requires expertise to identify characteristic waveform features and it is also prone to interobserver bias. As such, numerous studies have implemented and proposed deep learning algorithms for ECG analysis, which both eliminate these issues and reduce the time it takes to arrive at a classification, in place of manual analysis. However, these studies do not consider the importance of compression algorithms to condense ECG data into smaller, richer sets of data, which could significantly increase classification accuracy and significantly decrease processing time. Therefore, this study investigates 10 compressed formats—orthogonal leads, vectorcardiogram (VCG), three different PCA variations, three different median beat variations, autoencoder, and binary convolutional autoencoder (BCAE)—in comparison to the baseline 12-lead format in MI classification efficacy with the XResNet model.

When a person goes to check on a skin lesion or a runner wants to improve their form, a scanner is often used to track points on the body for measurements. Currently, there exists no solution that can instantaneously (in less than a second) compute the location of all these points at once. MiniMesh is a novel, resource-efficient algorithm that can accomplish this task on a small computer (like a phone or laptop) in real time from a single image. The algorithm takes an unorthodox approach of splitting this complex task into two simpler problems: finding the location of only 119 landmarks and extracting an outline from the patient’s image. After running these procedures, their output can be used to estimate the location of thousands of points which are displayed on the screen using a custom-made rendering engine. Overall, MiniMesh can process on average 20 images per second with high accuracy in all tasks. The speed of the algorithm can be improved to 50+ images per second when running each part of the algorithm parallelly. MiniMesh accomplishes what large motion capture systems can do using only a portable device, creating a fast, accurate, and inexpensive solution for all.

Climate change is the greatest existential threat to humanity. Throughout high school, I have connected myself to a massive global movement of youth fighting to change this. Over the past four years, I have dedicated my time to pursuing policy solutions to environmental issues and spearheading climate education initiatives in my local community, as well as at the national and global levels. Through these projects, I’ve engaged with other inspiring activists and learned the importance of collaboration and leadership - especially among young people whose futures are most impacted by global warming. When it came time for me to conduct my own research study in my senior year of high school, which I later submitted to Regeneron Science Talent Search, there was no question in my mind that I wanted to explore the social and political dimensions of climate change.

As Vladimir Putin retains his strong grip on political power in Russia, growing networks of conservative actors in the United States have drawn together the potential political futures of both countries (Stoeckl and Uzlaner, eds. 2020). Since the mid 2010s, conservative support for Russia has increased, a trend which continued during the commencement of the Ukrainian invasion in February 2022 (Gallup, 2022). Despite the bipartisan condemnation of the invasion and the humanitarian concerns expressed by many on the political left, several vocal leaders on the Christian right continue to express support for Putin and his authoritarian regime, and as data has demonstrated, this has significantly impacted the favorability that Americans show for Russia, especially the conservative far-right (Perry, Samuel L., 2021).

Recall that the problem statement of my project is to quickly and autonomously navigate and map a burning building, identify trapped inhabitants, and send recovery/escape routes to firefighters as fast as possible. During my background investigation of this subject, I discovered that navigating fires has traditionally posed a challenge for robotic systems. Traditional LiDAR (Light based) sensors are often employed for SLAM (Simultaneous Localization and Mapping: A class of perception and mapping algorithms), but these sensors lose accuracy significantly in smoke. Meanwhile, SONAR (Soundwave-based) or RADAR (Radio-based) mapping systems are often very bulky and technically complex, making them challenging to incorporate onto a dynamic robotic system. This posed a perception challenge: to develop a novel perception and navigation system that can function effectively in a fire environment.

Recall that the problem statement of my project is to quickly and autonomously navigate and map a burning building, identify trapped inhabitants, and send recovery/escape routes to firefighters as fast as possible. During my background investigation of this subject, I discovered that navigating fires has traditionally posed a challenge for robotic systems. Traditional LiDAR (Light based) sensors are often employed for SLAM (Simultaneous Localization and Mapping: A class of perception and mapping algorithms), but these sensors lose accuracy significantly in smoke. Meanwhile, SONAR (Soundwave-based) or RADAR (Radio-based) mapping systems are often very bulky and technically complex, making them challenging to incorporate onto a dynamic robotic system. This posed a perception challenge: to develop a novel perception and navigation system that can function effectively in a fire environment.