Chelsea Finn

Who she is

Chelsea Finn, born on 8 October 1992, is a computer scientist who works on machine learning and robotics. She is an assistant professor at Stanford University; her lab page describes her as an Assistant Professor in Computer Science and Electrical Engineering at Stanford and as a co-founder of Physical Intelligence^[3]. Her research includes meta-learning, deep reinforcement learning, and the joint learning of robot perception and control^[1].

Education

Finn studied electrical engineering and computer science at MIT as an undergraduate^[1]. She did her doctoral work at UC Berkeley and received her PhD there in 2018, advised by Pieter Abbeel and Sergey Levine^[1]. During her doctorate she interned at Google Brain, and she later worked at Google^[1].

Learning to learn

Much of Finn's work sits in meta-learning, and she is associated in particular with the MAML method, which is described as teaching systems how to learn how to learn^[1].

The idea is easiest to see against ordinary training. A conventional model is trained on one task — one dataset, one objective — and then applied. Change the task and you usually need another round of training with a large amount of new data. Meta-learning instead treats learning itself as the thing being trained: the system is exposed to many related tasks during training, and the goal is a model that can take on an unseen task from a small amount of data or a few steps of adaptation. In that framing, how quickly and how well the system adapts becomes the quantity being optimized, rather than performance on one fixed problem.

Why this matters is mostly about data. Many real settings do not come with a large dataset. A robot asked to handle an unfamiliar object, or a system asked to serve a new user, has only a handful of examples to work with. Whether a method can generalize from few examples is the difference between a system that must be retrained by an engineer and one that can adjust on its own. This is why learning to learn is usually discussed together with adaptation from few examples: the point is not merely to train faster in wall-clock terms, but to need less new information for each new task.

Robots, perception and control

Finn's other listed areas are deep reinforcement learning and the joint learning of robot perception and control^[1]. Reinforcement learning is the broad class of methods in which an agent improves a behavior by trial and error against a reward signal; deep reinforcement learning represents the policy or value function with a neural network rather than by hand. Joint learning of perception and control means that the part of a system turning sensor readings into an internal picture, and the part choosing actions, are trained together instead of being designed as separate modules, so that perception is shaped by what control actually needs. Put next to meta-learning, these areas point at the same ambition: a robot stack that can be given a new task and adapt, rather than being rewritten for every task.

Stanford and Physical Intelligence

Finn is an assistant professor at Stanford University^[1], and her lab lists the appointment as spanning computer science and electrical engineering^[3]. She is also a co-founder of Physical Intelligence and appears on the company's team page^[1][2].

Where her work sits

Her listed research areas — meta-learning, deep reinforcement learning, and robot learning — lie on the border between machine learning and physical systems. An algorithm that looks promising on a benchmark still has to survive real sensors, real actuation error, and thin data before it is useful on a machine. Working as a professor at Stanford and as a co-founder of a robotics company places her on both sides of that border^[1][2][3].