🧑‍🔬 Finale Doshi-Velez

Finale Doshi-Velez's papers here span two areas of machine learning. One is interpretability: what it means for a model's behaviour to be understandable to people, and how that quality might be assessed. The other is unsupervised learning of linguistic structure, using Bayesian models to recover syntax from plain text without annotated trees. Across both, the papers ask how a learning system can be kept within limits that make it useful, either to the humans who inspect it or to the cognitive constraints a model of language should respect.

The most cited work is Towards A Rigorous Science of Interpretable Machine Learning (2017), a position paper written with Been Kim. It offers no new models or datasets. Instead, it argues that the field lacks a shared definition or yardstick for interpretability and sets out a conceptual framework for one. That framework treats interpretability as the ability to explain a model's behaviour to a person, and it considers when such explanations are actually required, for example when goals like safety or fairness cannot be fully specified in advance. The second paper, Unsupervised Grammar Induction with Depth-bounded PCFG (2018), presents DB-PCFG. This Bayesian model learns a probabilistic context-free grammar directly and then converts it into a depth-bounded form. The bound reflects working-memory limits on center embedding and narrows the set of candidate parse trees.

The newer of the two papers is the 2018 grammar-induction study, which carries the computational-linguistics strand forward. It was written with Lifeng Jin, Timothy Miller, William Schuler and Lane Schwartz. The papers here date from 2017 and 2018 and connect most clearly to the topic of Interpretability.

Finale Doshi-Velez has 2 papers in our library, in 2017–2018, most of them on Interpretability; often with Been Kim, Lane Schwartz, Lifeng Jin.