Josh Gordon | Directed Study | 3 Credits
In this 3 credit 411 Directed Study, Josh Gordon will under the direction of Prof. Jeanna Matthews,
read and discuss research literature, collect and analyze data,
write research paper submissions and present his results verbally to a wider research team. The work
he is conducting and the related work he will be reading is in the high level areas of bias in natual language
processing and social media polarization. He will meet Tuesdays from 6-7 as well as frequently at other times
with Prof. Matthews and a wider research team.
He will produce a web portfolio of his research work. This course is not lecture based. Some of the work Josh does will be group work done in collaboration with other students.
Grading: 50% Participation and Verbal Presentation + 50% Web Portfolio of Research
Objectives: Conduct original research, gain skills in collection and anaysis of data including the production of
automated analysis scripts. Read and discuss related work.
Demonstrable Outcomes: Workshop and report on Generative Adversarial Networks for text processing
- Wrote and ran experiments to generate automatic defining sets
- Read(ing) papers on feature importance, GANs, machine learning bias, etc.
- Wrote part of a paper to submit to FAccT, did not submit paper
- Helped others write data gathering programs (Marzieh, Hunter, etc.)
- Meeting with Professor Mondal + Marzieh and working on modelling defining sets statistically
- Iterated through several models (manual, RANSAC, maximum-finding, now at GAN)
Papers read:
- Adversarial Text Generation Without Reinforcement Learning - https://arxiv.org/pdf/1810.06640.pdf
- TextKD-GAN: Text Generation using KnowledgeDistillation and Generative Adversarial Networks - https://arxiv.org/pdf/1905.01976.pdf
- Entropy and mutual information inmodels of deep neural networks - https://papers.nips.cc/paper/2018/file/6d0f846348a856321729a2f36734d1a7-Paper.pdf
- RelGAN: Relational Generative Adversarial Networks for Text Generation - https://openreview.net/pdf?id=rJedV3R5tm
- Adversarial Ranking for Language Generation - https://arxiv.org/pdf/1705.11001v3.pdf