Date: November 21, 2025
Title: ML for Networks in the Real World: Challenges Beyond Training the Models
Speaker: Jiayi (Jane) Chen, UT Austin
Abstract: There has been growing interest in applying machine learning to improve system performance, particularly because of its capability of adapting to heterogeneous and dynamic environments. But making ML effective in real systems requires addressing challenges that go well beyond training a model. In this talk, I will present two pieces of my work that tackle fundamental obstacles in this space.
The first challenge is accurate state representation. The performance of learned controllers depends heavily on the fidelity of their input features. Yet, when studying existing learned network controllers, we observe that they all rely on hand-made instantaneous or running-average metrics that provide coarse, delayed, or incomplete views of the true network state. This limits their ability to capture latent factors and adapt to non-stationary conditions. I will present UNUM, a unified network-state embedding framework based on Transformers, trained on diverse datasets to learn rich, latent representations of network behavior. I will show how UNUM can be integrated into existing learned and heuristic controllers, improving robustness and control performance across tasks.
The second challenge is efficient and adaptable deployment. For many system problems, constructing a single large learned model is impractical or unnecessarily costly. I will illustrate this through the design of CDN Hot Object Cache admission policies, where traffic patterns shift rapidly and objectives vary across deployments. I will present Darwin, a neural-aided expert selection system that dynamically chooses among a large set of candidate caching policies using a scalable bandit algorithm with side information. This approach enables flexible optimization of different caching objectives while maintaining low overhead and robustness to workload changes.
Bio: Jiayi (Jane) Chen is a fifth-year Ph.D. student in Computer Science at UT Austin, co-advised by Aditya Akella and Sanjay Shakkottai. Her research sits at the intersection of machine learning, networking, and systems, with a focus on using learned models to improve the performance and robustness of networked systems. She is also interested in designing system interfaces and abstractions that make it easier to integrate learned components into real-world systems. She is a part of the Learning-Directed Operating Systems (LDOS) Expedition.