Talks and Presentations

I am honored to give talks at the following conferences and engage with a broader community of scholars. The presentation lists are in reverse chronological order.

1. Deep Galerkin Method for Mean-Field Control Problem

Jingruo Sun, [Poster], INFORMS 2024 Annual Conference

We consider a mean-field control problem focusing on the average welfare of weakly interacting agents under finite-state space, continuous-time, and finite horizon. The value function is characterized as the unique viscosity solution of a Hamilton-Jacobi-Bellman (HJB) equation in the simplex. We construct a Deep Galerkin Method (DGM) to solve the HJB equation. DGM approximates the solution using a deep neural network which is trained to satisfy the differential operator and boundary conditions of the high-dimensional nonlinear PDE. The accuracy of DGM is validated through various numerical experiments. As the theoretical support, we prove that the loss function converges to zero, and the corresponding neural network approximators converge uniformly to the true value function on the simplex. Therefore, we propose a comprehensive framework for the high-dimensional mean-field control problem, leading to a reliable and efficient methodology for decision-making.

2. The Development of Facticity: from Preliminary Findings to Accepted Implicit Knowledge

Tianyu Du, Jingruo Sun, and Yuze Sui, [Poster], IC2S2 2024, ICSSI 2024

This paper examines how scientific ideas transition from novel claims in published research to established, implicitly accepted facts. We propose a multi-phase process: (1) only a fraction of newly introduced ideas, initially shared via preprint and receive attention and citations; (2) those ideas acquire stable associations, lose citation salience, and become categorical concepts, signaling emerging consensus; and (3) eventually, those ideas appear in reference materials such as textbooks or Wikipedia, indicating fully integrated knowledge. This progression can reverse when previously accepted facts are contested, moving them back to the “bleeding edge” of research. Using data from arXiv, Web of Science, and Wikipedia, we trace how focal concepts move through these phases—analyzing citation patterns, co-citation networks, and whether concepts are mentioned but not formally cited. We illustrate this process by charting the life trajectory of “Ricci flow” in mathematics, revealing how key ideas become consolidated over time.

3. A Greener Future Beyond Profits: Sustainability as a Driver of Market

Jingruo Sun, departmental talk [Poster], Best Project Award of Machine Learning (1/172), 2023

Sustainability has become pivotal in decision-making, with Environment, Social, and Governance (ESG) factors emerging as critical indicators of responsible, long-term organizational behavior. This project explores how ESG metrics reveal patterns of sustainable performance using advanced machine-learning techniques to handle complex, high-dimensional data. We employ predictive modeling, feature selection, and multi-objective optimization to assess the relationship between ESG indicators and long-term performance, balancing model complexity, interpretability, and reliability. We provide a robust framework for evaluating ESG variables as actionable, data-driven insights for guiding responsible decision-making and promoting the long-term positive impact of sustainability.

4. Importance-Weighted Sampling Enhanced VAE for MIRT Model

Jingruo Sun, departmental talk, Undergraduate Honors Thesis for Statistics, 2023

Multidimensional Item Response Theory provides an ideal foundation for modeling performance in complex domains, taking into account multiple basic abilities simultaneously, and representing different mixtures of the abilities required for different test items. However, with the increasing size of modern assessment data, conventional estimation methods become computationally demanding, and hence they are not scalable to big data. To tackle this challenge, we present a novel approach utilizing importance-weighted sampling enhanced Variational Autoencoder on logistic models. We leverage the power of variational inference from the field of machine learning to effectively approximate the elusive marginal likelihood. We further enhance our method by using importance-weighted samples to yield a superior log-likelihood approximation.