Responsible AI and Human Centred Technology

Mentors: Student: Taylor Lee

Artificial intelligence developments can lead to growth and scientific advancement at a rapid pace making it even more important how these systems are implemented and informed to the general public in the most accessible manner. Data Cards Playbook is about bringing people together to transparently show the importance of transparency, ethics, and dataset literacy. The goal is to translate the playbook activity to be designed with users in a participatory design setting as hoped in the ACM Track 3 proposal. This is where interactive experiences on the website can add even more magic to what can be an even more powerful tool. The goal is to turn one of the most essential, popular static activities as determined by the team and users into something interactive. This would include user research, heuristic evaluations, wireframing and designing UIs, and prototyping the product in code. Ultimately, the hope is to bring comprehensive transparency artifacts that are not only beneficial to the documentation process and stakeholders themselves but to the human-centric and responsible practices when it comes to datasets in ML applications.

Mentors: Student: Chris Schmitz

This project seeks to experimentally examine and benchmark the performance of fair data reweighting (FDW), a fairness preprocessing technique, as part of TensorFlow’s model remediation library. The approach is a set of experiments which evaluates the impact of FDW on both traditional model metrics like accuracy or R2, and on different fairness metrics (prediction disparity between groups; accuracy between groups), across different datasets, machine learning models, and hyperparameter configurations. This is then benchmarked against other fairness remediation techniques such as mindiff across the same indicators. Further, documentation is created to ease understanding and use of FDW. This includes a text tutorial to be included on TensorFlow's Responsible AI site, and colab notebooks demonstrating how to use FDW on demo datasets. The minimum deliverables are the insights gained by the experiments in the form of a report, and a set of documentation, both in the form of tutorials and colab code samples, showing how to use FDW. Optional additional deliverables may include changes to the FDW implementation itself or further integration into the TF ecosystem.

Mentors: Student: Anjishnu Mukherjee

This project focuses on providing easy-to-understand walk-throughs of standard workflows for analyzing models using LIT and also providing in-depth how-tos for using and building on top of some of the LIT modules which are relevant for these workflows. Providing this content in the form of tutorials and code contributions over the summer, I hope to increase the accessibility to the tool for different types of users.

Mentors: Student: Aryan Chaurasia

This project is about adding new AI models for LIT demos. The whole idea is to include different kinds of models as examples, so people from various backgrounds can easily see and understand model demos and learn to use LIT for their own custom models. I will be focusing on two models: a multilingual question answering model based on the TyDiQA dataset, and another model that generates images from text (the Dalle Mini model).