Underwater instance segmentation models are typically trained and evaluated on a single dataset, leaving cross-dataset generalization unstudied. This project measures the cross-dataset domain gap between TrashCan (7,212 deep-sea ROV images, 22 classes) and SeaClear (8,610 shallow-water images, 40 classes) using Mask R-CNN with a COCO-pretrained ResNet-50 FPN backbone. Models trained on one dataset fail catastrophically on the other despite overlapping object categories. We show this gap is visual, not semantic: silhouette analysis of backbone features reveals strong dataset clustering even at the per-class level. To mitigate the gap, we pool both datasets into common category spaces via a generic coarsening hierarchy and show that a single pooled model recovers or exceeds in-domain performance on both test sets. We also re-split TrashCan using frame-chunking to fix data leakage in the original split. All results are explorable via an interactive Streamlit dashboard.

Question:
What early signals predict sustained success in the music industry?

Objective:
Many musicians produce hit songs, but not all are able to do so more than once. This project builds a machine learning classifier to distinguish hitmakers (artists with multiple top 20 Billboard Hot 100 hits) from one-hit wonders, using only information available at the moment of a musician’s first top 20 hit song.

Conclusions:
Our model reveals that prior charting experience, collaboration network position, chart longevity, genre breadth, and dominant genre affiliations are the strongest predictors of sustained success.

Data sources:
- MusicBrainz (artist metadata, genre tags, collaboration graph)
- Billboard Hot 100 & 200 chart data
- Spotify (artist and song metadata)
- Google Trends (relative search volume at time of first hit song)

A polyp is an abnormal tissue growth in the large intestine that is typically benign but can develop into malignant colorectal cancer. Colonoscopy enables endoscopists to identify and assess these polyps for potential removal. However, the accuracy of this procedure depends heavily on the clinician’s expertise, making it prone to human error and variability. Our goal is to build a deep-learning model that detects colorectal polyps in images from colonoscopies to minimize missed lesions and improve patient outcomes.

Lead is a toxic metal, and in children especially, lead exposure can have severe health consequences -- even small amounts of lead have the potential to affect memory, behavior, and learning ability. Despite this, numerous schools across New York State have at least one drinking water outlet with lead levels testing for above 5 ppb. In this project, we aim to predict the presence of lead contamination in school drinking water, and better understand the role of demographic, socioeconomic, infrastructural, and geographic features in elevated lead levels.

Fraud detection is a critical area where deep learning has been effectively applied to identify and prevent unauthorized transactions, money laundering, and other financial crimes. Traditional rule-based systems and statistical models often struggle to detect sophisticated fraud patterns, particularly when dealing with large volumes of data and rapidly evolving fraud techniques. In contrast, deep learning models, such as CNNs, RNNs, and autoencoders, have proven highly effective in analyzing complex, high-dimensional transaction data and detecting subtle, non-linear patterns indicative of fraudulent activity.
In this project, we build a User ID-based fraud detection model using autoencoders, trained on unlabelled real-world credit card transaction data, capable of detecting fraud with a precision of up to 35% and a recall of up to 72%, performing significantly better than traditional ML/statistical baseline models..