Machine Learning / Artificial Intelligence Seminar - Mahsa Bastankhah
October 2, 2026 10:00AM—11:00AM
Location:
In Person
-
ASA Conference Room, Gates Hillman 6115
Speaker:
MAHSA BASTANKHAH,
Ph.D. Student, Department of Electrical and Computer Engineering, Princeton University
https://sites.google.com/view/mahsabastankhah/home
Supervised learning has achieved tremendous success in extracting patterns from large datasets. However, it remains unclear whether the same models and data can help solve novel tasks in data-scarce regimes, where massive datasets cannot be collected, and intelligent data collection through interaction with the environment (i.e., RL) is itself a big part of the problem. Efficient data collection requires the agent to learn an abstract representation of the environment that does not attend to every detail, but instead captures the core factors governing the dynamics. Otherwise, the real world is so complex that no agent can model it (the only computer that can model it is the real world itself). In this work, we argue that representation learning should be intertwined with learning to control the environment through interaction. We show that empowerment –the channel capacity between an agent’s actions and future states– that has been historically used for diverse behavior learning in RL, is a principled objective for learning these representations. Maximizing this mutual information requires the agent to discover different axes of controllability in the environment. Our main contribution is to show that empowerment agents learn two distinct and complementary representations of the state (forward and backward representations), both centered on control and both discarding control-irrelevant factors, and that these representations substantially improve performance in complex environments with high-dimensional distractors. Overall, our work establishes empowerment as a principled objective for learning intrinsic representations of an MDP that are independent of any particular dataset or downstream task, and are centered on control—the defining aspect that distinguishes reinforcement learning from supervised learning.
—
Mahsa Bastankhah is a fourth-year PhD student at Princeton. She is advised by Prof. Benjamin Eysenbach. Her work is focused on exploration and skill learning in unsupervised RL. She is interested in emergent exploration and understanding how agents learn to do complex tasks without hand-designed curricula and what the key ingredients of such exploration are. She is also interested in abstract representation learning and understanding the interplay between representation learning and exploration in high-dimensional environments.
For More Information:
nihars@cs.cmu.edu