AIAI Seminar-Monday 29 April-Talk by Dilara Kekulluoglu and Xenia Heilmann

Speaker: Dilara Kekulluoglu

Title: Answerable Sociotechnical Systems

Abstract: Sociotechnical systems (STSs) combine people and machines to take actions. Artificial intelligence (AI) enables STS to make increasingly autonomous decisions that impact human lives. Their reasoning processes still often remain unclear to people interacting with such systems, which may also harm people by making unjust decisions. There are no efficient means for people to challenge automated decisions and obtain proper restitution if necessary. On the other hand, organizations may be willing to provide more transparency about their decision-making process, but answering each of the questions people ask could be cumbersome. It is also not always clear who is qualified and accountable to answer to the people harmed by autonomous decisions. We propose a mediator agent framework that will bridge the gap between organizations that employ AI and people who were harmed by the automated decisions. Our approach helps the organizations to implement more answerable AI practices, and it also empowers people to ask for clarifications, request updates on actions as well as remedies through dialogues.

Speaker: Xenia Heilmann

Title: Differentially Private Sum-Product Networks

Abstract:  Differentially private ML approaches seek to learn models which may be publicly released while guaranteeing that the input data is kept private. One issue with this construction is that further model releases based on the same training data (e.g. for a new task) incur a further privacy budget cost. Privacy-preserving synthetic data generation is one possible solution to this conundrum. However, models trained on synthetic private data struggle to approach the performance of private, ad-hoc models. In this talk, I will give a short introduction to differential privacy and present a method based on sum-product networks that is able to perform both privacy-preserving classification and privacy-preserving data generation with a single model.