-
Email Content Extraction
Proposed a double fine-tuning method to train a model that can extract key information from promotional emails from student organizations.
-
Infinite-Width 1-Layer ReLU Networks with L2 Regularization on 2D Data
Given a dataset of two-dimensional input and one-dimensional output, investigated the set of 1-layer ReLU networks that interpolate the dataset and, among such interpolants, minimize the L2 norm of the weights. Considered datasets where the points either form the vertices of a regular polygon, or are symmetric with respect to a line.
-
What Makes In-Context Learning Work On Generative QA Tasks?
Analyzed the effect of different components of in-context learning demonstrations on the capabilities of large language models on generative QA tasks.
-
Data Valuation with Shapley Values
Analyzed the robustness of Shapley values across different training settings. Proposed a novel approach of approximating Shapley values by evaluating on simpler models with similarly expressive power
-
PhoLSTM - A Joint Multi-task Model in Vietnamese Natural Language Processing
Proposed a variation of a joint multi-task language model for Vietnamese that outperforms the baseline model
-
Extension of Simple Algorithms to the Matroid Secretary Problem
Proposed multiple algorithms for the Matroid Secretary Problem generalized from the ordinary Secretary Problem. Proved their properties using the greedy algorithm and forbidden set frameworks
-
Free-Order Matroid Secretary Problem
Proposed a variation to an algorithm with a competitive ratio of 4. For uniform matroids, proved that the variation has a constant competitive ratio. Conjectured that the ratio converges to e if the size of the matroid tends to infinity. For general matroids, empirically observed that the algorithm is likely to have a competitive ratio close to e
-
Applications of Polynomial Methods
Studied how polynomial methods can be applied to Joints and Nikodym Problems
-
Prediction of Seoul Metro Ridership
Designed a LSTM model to predict the hourly ridership of Seoul metro
-
Resiliency of East Asian Air Network to Adversarial Attacks
Through simulations, presented that the East Asian air network is more robust to adversarial attacks than the global network