# Chelsea Finn > ~1989– · Computer Scientist, AI Researcher > > **Recorded contribution:** Meta-learning; MAML; robot learning; Stanford ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Computer scientist Chelsea Finn is a Stanford professor working on robot learning, reinforcement learning, and meta-learning. She led the 2017 model-agnostic meta-learning (MAML) paper, which learns an initialization from which a model can adapt to a new task with a small number of gradient steps. Her research also studies visual foresight, imitation, and generalist robotic behavior. MAML searches for parameters that are not merely good on training tasks but are positioned for rapid improvement on a new task. The outer loop evaluates how well a small inner-loop update performs, then differentiates through that adaptation process to improve the shared initialization. ## 2. The problem inherited A separately trained model for every task requires too much data and interaction, while robots need to adapt quickly when objects, goals, or environments change. ## 3. The central contribution Finn and collaborators introduced MAML, a general optimization approach for learning parameters explicitly chosen for rapid gradient-based adaptation. ## 4. Reconstruct the mechanism 1. Sample a task and split its examples into adaptation and evaluation sets. 2. Take one or more inner-loop gradient steps from shared initial parameters on the adaptation examples. 3. Measure the adapted model on the task's evaluation examples. 4. Differentiate that post-adaptation loss through the inner update to improve the shared initialization. ## 5. What changed downstream - MAML became a foundational meta-learning algorithm across supervised, reinforcement, and imitation learning. - It sharpened the idea that the learned object can be an adaptation procedure, not only a fixed predictor. - The method became a reference point for gradient-based meta-learning and influenced few-shot classification, robot adaptation, and analysis of what it means for a system to learn how to learn. ## 6. Attribution, limits, and uncertainty - MAML has three authors and builds on hierarchical learning, multitask learning, and earlier meta-learning. - Second-order optimization can be expensive, task distributions may be unrealistic, and fast adaptation within one family is not general intelligence. - Differentiating through adaptation is expensive, performance depends on the task distribution, and fast optimization on benchmark tasks does not guarantee safe transfer to genuinely novel physical conditions. ## 7. Reconstruction lab Meta-train a linear regressor across sine-wave tasks. Compare random and learned initialization after one gradient step and test a frequency outside the training distribution. Train on a narrow and broad task distribution, then evaluate an out-of-family task and show whether rapid loss reduction reflects reusable structure or memorized similarity. ## 8. Evidence trail - [Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks](https://proceedings.mlr.press/v70/finn17a.html) — ICML / PMLR - [Chelsea Finn](https://ai.stanford.edu/~cbfinn/) — Stanford University - [Chelsea Finn](https://en.wikipedia.org/wiki/Chelsea_Finn) — Wikipedia contributors --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*