The CDS Industry Mentorship Program is an exclusive benefit for members of the Industry Affiliates Program. The mentorship program matches small teams of data science master’s students with an industry-proposed project. Over the course of an academic semester, each team works under the guidance of an industry mentor, And many past teams have had papers accepted at prestigious conferences, such as EMNLP. This year’s cohort consisted of over 70 students working on 22 projects with Amazon, Chan Zuckerberg Initiative, Facebook, GE Healthcare, Goldman Sachs, IBM, Lexalytics, Microsoft, Oracle, Raytheon, Stanley Black & Decker, the UMass Classics department, and Voya Financial. Two of the teams have had papers accepted at conferences or workshops this year.
One such team member, Pulkit Sharma, a second-year master’s student, participated in the Industry Mentorship course (COMPSCI 696DS) to gain experience working directly with data science professionals in industry. His team consisted of two other master’s students, Apurva Bhandari and Shezan Rohinton Mirzan, and four Microsoft representatives, two of whom are alum of UMass Amherst CICS. Their project evaluated methods for explaining machine learning models on factors such as computation time, significance of attribution value, and explanation accuracy. Despite the interruption of the COVID-19 pandemic, the students continued their project, working remotely with their industry mentors.
“The mentors were very responsive to our technical questions and provided insights on learning how to do research in general,” Sharma said. “I learned a lot from them and also got experience in the process of writing a paper.” The team’s paper, "Evaluating Tree Explanation Methods for Anomaly Reasoning: A Case Study of SHAP TreeExplainer and TreeInterpreter,” has been accepted to CMAI 2020 (Conceptual Modeling Meets Artificial Intelligence and Data-Driven Decision Making).
Master’s students Wenlong Zhao, Saurabh Kumar Shashidhar, Prafull Prakash were assigned to one of the three Amazon Alexa teams in place this year, working on methods for sparsification of word embeddings, along with three Amazon mentors. The team produced a short paper titled “Compressing Transformer-Based Semantic Parsing Models using Compositional Code Embeddings,” which describes a process to reduce the size of parsing models for deployment on edge devices with limited memory, such as Amazon Alexa or Google Assistant devices. Their paper was recently accepted at EMNLP 2020.