[Seminar] MLDS Unit Seminar 2024-8 by Ms. Weiqiu You (University of Pennsylvania) , Mr. Xiaoyuan Zhang (City University of Hong Kong), at D23, Lab5

Speaker 1: Ms. Weiqiu You (University of Pennsylvania)

Title: Sum-of-Parts: Faithful Attributions for Groups of Features

Abstract: An explanation of a machine learning model is considered "faithful" if it accurately reflects the model's decision-making process. However, explanations such as feature attributions for deep learning are not guaranteed to be faithful, and can produce potentially misleading interpretations. On the other hand, models designed with provably faithful attributions often come with a severe trade-off in performance. In this talk, I'll present Sum-of-Parts (SOP), a framework that transforms any differentiable model into a faithful model whose predictions can be attributed to groups of features. The SOP framework can use powerful deep learning backbones and be trained end-to-end, capable of automatically learning groups of features useful for prediction without direct supervision. With these capabilities, the SOP framework achieves high performance on vision benchmarks, and generates more human interpretable groups than other faithful feature attribution methods. In a case study, we use the faithful group explanations from SOP to help cosmologists discover new knowledge about galaxy formation.

Speaker 2:  Mr. Xiaoyuan Zhang (City University of Hong Kong)

Title :  Multiobjective meets machine learning, from a single solution, a finite set of solutions to infinite solutions. 

Abstract:   Multiobjective optimization is prevalent in machine learning applications like fairness classification, domain adaptation, and multitask learning. A current hot topic is multiobjective machine learning (MOML), which applies multiobjective optimization perspectives to machine learning.

Recent work in MOML falls into three categories: finding a single solution, finding a set of solutions, and finding an infinite number of solutions. In this talk, I will briefly review significant recent works in each category and introduce my contributions.

For finding a single Pareto solution, I will present PMGDA, a preference-based multiple gradient descent algorithm. For finding a diverse set of solutions, I will discuss maximizing the minimal pairwise Pareto objectives on the Pareto front. Finally, I will introduce our latest work on finding infinite Pareto solutions using a neural model.