JavaScript seems to be disabled on this browser. My website does not function properly without it!
Machine Learning
Assertion Extraction
The objective of this project was to train a machine learning model to dig the web and extract assertions made by well-defined sources. Assertions were considered triples of the form {source, manner, message}. For example, {Michelle Obama, Wrote, People want to feel hopeful} is a valid assertion. Stanford CoreNLP tools were used to process a small annotated corpus, and features were hand-engineered and used in a supervised ML setting. A Support Vector Machine model was trained and tuned using the SVMLight tool. The results showed promising improvements over the baseline of manually designed rules (precision/recall scores of 94.44/50.90), but further improvement was limited to the syntactic parser performance. [...]
Deep Learning
Deep Learning Effectiveness for Scheduling
We investigated the efficacy of depth in neural networks when applied to a classification task for the Job-Shop Scheduling Problem. Both hand-engineered features and raw features were used to assess the ability of deep networks in both combination and construction of high-level features. The label is a standardized form of the difference between the optimal makespan of an instance with the average of same-size instances. The results indicated improvements over the shallow baseline for both sets of features, but more remarkably when raw features were used. The important conclusion was that neural networks can outperform conventional supervised machine learning algorithms for a problem with complex correlations between the features and the label. [...]