CS Student Seminar
University of Cincinnati
Spring 2024

Host: Vikram Ravindra and Tianyu Jiang
Location: Baldwin 749
Time: 3:30 - 4:30 pm


Welcome to the homepage of computer science student seminar! We will be happy to have any students/faculties, CS/non-CS join us!

This seminar is mainly for Ph.D. students, postdocs or thesis-based M.S. students in the department of Computer Science to share their work, get new ideas from discussions and practice presentation skills.

UC CS Ph.D. students: you are encouraged to present your ongoing research, have a practice talk for your upcoming defense or conference oral presentation, give a tutorial on interesting tools or new techniques, and so on. Each session consists of a 40-50 minutes talk followed by Q&A. Please don't hesitate to email Dr. Ravindra or me to register as a speaker!

Date Speaker Title
02/12 Usman Anjum Model-Agnostic Meta-Learning for Task Adaptation to Few-shot Learning
Abstract: Meta-Learning (or "learning-to-learn") and few-shot learning is the task of learning from one set of data and using the information from a small sample of data to generalize to an unseen data (or data that was absent in the training set). This can be done by either understanding the similarity between the datasets or using the parameters of the training model. In this presentation, we will explore a class of meta-learning algorithms that perform small gradient updates to the parameters of a model so that the model can adapt to the new task and achieve good generalization performance.

Bio: Dr. Anjum is a postdoc researcher working with Justin Zhan in the CS department of UC.
02/26 Vikram Suresh SAT and Student Aptitude Decline, Evidence from an AI Experiment
Abstract: Utilizing first-of-a-kind "Transformed Control" method using OpenAI's GPT-4 model, this paper investigates the changes in SAT Math section difficulty and its reflection on high school students' math proficiency from 2008 to 2023. Our analysis reveals a noticeable decline in the rigor of SAT Math questions over time, culminating in the GPT-4 model scoring an average of 71 points higher in 2023 compared to 2008. This reveals a steeper 99-point decline in high school student proficiency between 2008 and 2023. In an era where universities are reevaluating the relevance of SAT scores for admissions, our results illuminate a twofold trend: a decrease in the mathematical challenge posed by the SAT and a concurrent decline in students' mathematical aptitude, signaling significant shifts in educational assessments.

Bio: Dr. Suresh is a postdoc researcher working with Hans Breiter in the CS department of UC.
03/18 Sina Eghbal A Brief Introduction to Information Theory
Abstract: Information theory is the study of quantifying, storing, and communicating information. This presentation gives an introduction to information theory. We start by discusses information, its definition, information sources, and how information is modelled and quantified. We then proceed to explain and provide an intuitive understanding of basic definitions, including entropy, joint entropy, and conditional entropy, and examine the information in a few simple data sources. Finally, we discuss compression as a common application of information theory, looking at Huffman coding and how it can be used to achieve epsilon-optimal compression for any data source.
Bio: Sina is a phd student working with Badri Vellambi in the ECE department of UC.
04/15 Shemon Rawat PEDS: Ask for Price and Explanations for Data Sharing
Abstract: Data sharing and trading are integral to modern society, yet current marketplace services often lack data pricing models that use data itself to evaluate the price of the data, leading to inefficiencies and over-purchasing of data. Existing pricing methods, such as fixed view rates or query-based pricing, have limitations in flexibility and fairness. In this paper, we introduce PEDS (Price and Explanations for Data Sharing), a system designed to address these issues. PEDS calculates the price of data based on the Information Gain (IG) generated by integrating data. By incorporating IG into pricing, PEDS offers a flexible pricing structure tailored to the unique insights each query offers. Additionally, PEDS provides a method for generating meaningful summaries that explain the factors contributing to the estimated price. Our methodology provides valuable insights into the implications of shared data, offering a practical solution for understanding and harnessing the benefits of data sharing in modern environments.
Bio: Shemon Rawat is a phd student working with Seokki Lee in the CS department of UC.