Alyssa Herbst, DAC M.S. student in the Department of Computer Science

Graphic is from Herbst’s research on “Active Learning by Greedy Split and Label Exploration”

After receiving a master’s degree in computer science at Fall Commencement, Alyssa Herbst will head to New York City. She has already accepted a position as software engineer at Instagram, where she interned on the Shopping Machine Learning team this past summer.

Herbst’s interest in machine learning sparked when, as an undergrad in the Department of Computer Science, she took a class taught by Bert Huang. She wound up working in Huang’s Machine Learning Laboratory on a twitter scraping project to assist with cyberbullying research.

“As part of this research, we had a corpus of tweets that we wanted to label as either ‘bullying’ or ‘not bullying,’ but a limited crowdsourcing budget. So we started to think about what it would look like to ‘guess’ the labels of tweets with some degree of certainty if crowdsource workers labeled some of the tweets,” said Herbst.

“I came into the lab with little knowledge about machine learning and ended up learning so much from attending meetings and going over research papers,” said Herbst. “I was happy to be able to develop my own research as a grad student with Professor Huang.” Huang is now her advisor at the Discovery Analytics Center.

Herbst said she really appreciates the opportunities she has had at DAC  “to collaborate with really smart, talented people and to gain exposure to other areas of research.”

Her current work, a form of active learning, or human-in-the-loop machine learning, is an extension of the twitter scraping project she worked on as an undergrad.

“We aim to guess the labels of a large unlabeled dataset in a low-budget setting by having a person iteratively label data from the dataset and split apart the data into groups when we see that label groupings have emerged,” she said.

“Data with similar features to a label grouping will likely have the same label. We make decisions on when to split data into groups and when to label data by using bounds that tell us how confident we are about the labels we infer,” said Herbst.

She and Huang have collaborated on “Active Learning by Greedy Split and Label Exploration.”  And last December, she presented “Interactive Learning by Uniformity Propagation Strategy” at the Women in Machine Learning poster session at the Conference on Neural Information Processing Systems (NeurIPS 2018).