Computer Science (CS) courses

(CS221) Newspaper GAL: Newspaper Generation and Analysis of Leaning – with S. Bea and L. Neustock

The aim of this project is twofold. Firstly, we detect bias in newspaper articles and in their corresponding the Facebook comments by applying sentiment analysis. The results confirm different leanings in newspaper sources. Secondly, the dataset is used to create a comprehensive “Must-Read” list for a news topic incorporating different viewpoints, achieving greater coverage of opinions while remaining overall neutral.

(CS231N) Enhancing Prototypical Networks for Within- and Cross-Domain Few-Shot Learning – with X. Chen and Y. Ye

The goal of few-shot learning is to train a classifier for distinguishing images from novel classes based on a limited amount of labeled examples. Many existing studies on few-shot learning assume the within-domain setting, in which examples from base and novel classes come from the same domain (e.g. hand-written characters). A related but more challenging problem is cross-domain few-shot learning, in which there is large domain discrepancy between base and novel classes. In this project, we tackle both problems by 1) fine-tuning Prototypical Networks with a pre-trained EfficientNet as the backbone and 2) applying a data augmentation technique called mixup. Experimental results demonstrate that our approach attains state-of-the-art performance on 5-way 5-shot learning tasks evaluated on CIFAR100. Additionally we achieve reasonable results in experiments of miniImagenet comparing with the work of matching network and ProtoNet.