This project enhances cross-lingual information retrieval (CIR) for low-resource languages through two main strategies: query augmentation, which generates diverse queries using large language models to enrich training data, and a weighted loss function that prioritizes underrepresented languages for balanced model training. Our methods significantly improve retrieval quality, showcasing the potential for effective CIR systems even in the face of limited resources.

This study investigates how the sequence of training environments affects performance in simple driving tasks for an autonomous driving agent. By training agents solely through interaction with maps of varying difficulty, we demonstrate that transfer learning enhances performance within single-environment driving scenarios. However, we find that agents struggle to master advanced driving capabilities and fail to generalize well to new environments, regardless of the sequence of training data. We conclude by looking at areas to build on this work such by combining imitation learning with curriculum learning and developing curriculum-specific MDP.

This study examines the causal impact of being placed on the Dean’s List, a positive education incentive, on future student performance using a regression discontinuity design. The results suggest that for students with low prior academic performance and who are native English speakers, there is a positive impact of being on the Dean’s List on the probability of getting onto the Dean’s List in the following year. However, being on the Dean’s List does not appear to have a statistically significant effect on subsequent GPA, total credits taken, dropout rates, or the probability of graduating within four years. These findings suggest that a place on the Dean’s List may not be a strong motivator for students to improve their academic performance and achieve better outcomes.

This study explores machine learning algorithms to predict movie ratings (positive or negative) using 2,500 IMDB reviews. The data was preprocessed into a TF-IDF matrix with 17,000 features, reduced via PCA and heuristic term selection. Logistic regression with gradient descent achieved a 20% test error rate (5.35% cross-validation error) using 1,000 terms. K-nearest neighbors (K-NN) reached a 26.2% test error rate (23.7% cross-validation error) with 100 terms and k=8. Logistic regression outperformed K-NN in accuracy but required more computational resources. The findings highlight trade-offs between accuracy and efficiency in text classification, suggesting potential for advanced methods like neural networks or support vector machines.

This study applies Quadratic Discriminant Analysis (QDA) to predict diabetes status using the Pima Indians Diabetes dataset, comprising 392 observations and 8 predictors. Exploratory analysis revealed significant covariance differences between diabetes-positive and diabetes-negative groups , supporting the use of QDA over Linear Discriminant Analysis. QDA achieved an average accuracy of 79% with a 70-30 train-test split and 77.69% via 5-fold cross-validation. Principal Component Analysis (PCA) was used to reduce dimensionality, with the five-component model achieving 78.5% accuracy while retaining 80% of explained variance. The study demonstrates QDA’s utility in diabetes classification, with PCA balancing model complexity and performance.