Esther Rolf Personal Website

I study statistical and geospatial machine learning. My research blends methodological and applied techniques to study and design machine learning algorithms and systems with an emphasis on usability, data-efficiency and fairness. My current research directions include developing algorithms and infrastructure for reliable environmental monitoring using machine learning, and understanding the multifaceted nature of representation in data and how that affects our ability to train fair and effective machine learning systems.

In fall of 2024 I’ll join CU Boulder as an assistant professor of computer science. See below for information on joining my lab at CU Boulder!

I’m currently a visiting professor at CU Boulder and a postdoctoral fellow at Harvard with the Harvard Data Science Initiative (HDSI) and Center for Research on Computation and Society (CRCS).

Research

In my research I develop and analyze of machine learning algorithms to address pressing social and environmental problems. Sometimes this entails developing or analyzing new statistical ML techniques, sometimes this entails carefully applying, adapting, and evaluating ML methods in a specific context of use; most often it entails a blend of the two. Some specific project areas and application contexts:

Tailored machine learning for remotely sensed data (e.g. using satellite imagery for environmental monitoring).
Characterizing and formalizing notions of representivity in training data (e.g. numerical representation: how many data points come from each source or group?, as distinct from what components of an individual or environment does a collection of data actually reflect?), and how these notions of representation effect our ability to train fair and useful machine learning systems.
Understanding and addressing key challenges in geospatial machine learning (e.g. spatial error structures make it hard to evaluate geospatial ML models, and can introduce concerns of bias or unfairness in downstream use).

For a full list of papers please see my google scholar page.

Selected Projects

New! "Geographic location encoding with spherical harmonics and sinusoidal representation networks" (w/ Marc Rußwurm, Konstantin Klemmer, Robin Zbinden, and Devis Tuia) will appear at ICLR 2024.

"Fairness and representation in satellite-based poverty maps: Evidence of urban-rural disparities and their impacts on downstream policy" (w/ Emily Aiken and Joshua Blumenstock) appeared at IJCAI 2023.

Figure showing key distinctions of geospatial ML from other learning paradigms: data are often autocorrelated and only available in certain locations, while prediction models are often expected to geographically extrapolate (diagram showing this: train in one region, use in another region ).

Evaluation Challenges for Geospatial ML

Esther Rolf. Workshop on Machine Learning for Remote Sensing at ICLR, 2023.

Figure showing that using implicit posterior models, we can resolve coarse landcover labels into high-resolution landcover predictions, by pairing uncertain prior beliefs based on avilable landcover data with high resolution satellite image observations.

Resolving label uncertainty with implicit posterior models

Esther Rolf*, Nikolay Malkin*, Alexandros Graikos, Ana Jojic, Caleb Robinson, Nebojsa Jojic. UAI, 2022.

Oral presentation (5% of accepted papers)

$Figure illustrating that by parametrizing the training dataset in terms of the fraction of training samples from each of many population groups, we can calculate the optimal training distribution for learning procedures under different conditions. Doing so formalizes the important of sub-group representation of small groups in training data, even when the objective is population average accuracy.$

Representation matters: Assessing the importance of subgroup allocations in training data

Esther Rolf, Theodora Worledge, Benjamin Recht, Michael I. Jordan. ICML, 2021.

A schematic of the MOSAIKS system for satellite imagery and machine learning. Satellite images are converted into a fixed embedding matrix X, which can be used to solve downstream prediction tasks with a simple linear regression model (shown here are predictions for forest cover and population density).

A generalizable and accessible approach to machine learning with global satellite imagery

Esther Rolf*, Jonathan Proctor*, Tamma Carleton*, Ian Bolliger*, Vaishaal Shankar*, Miyabi Ishihara, Benjamin Recht, Solomon Hsiang. Nature Communications, 2021.

How to join!

I am looking for postdocs and PhD Students to join my lab at CU, starting as soon as the fall of 2024, who are interested in

statistical and/or geospatial ML methodology, and
context-driven research for real-world problems

and who want to work in and foster a lab environment which is

doing innovative, sometimes interdisciplinary research
collaborative, supportive, and communication-forward.

Interested in a PhD? Apply to CU Boulder's PhD program in computer science and list me as a potential advisor. Applications open in the fall for positions in fall of the following year.

Interested in a postdoc? Send me an email at esther.rolf@colorado.edu describing your background, research interests, and fit with my lab.

Other? Email me at esther.rolf@colorado.edu. Please note that I will not be able to respond to all emails.

Courses

Current Topics in Computer Science: Geospatial and Statistical Machine Learning

Fall 2024 Tues/Thurs 12:30-1:45pm, CU Boulder (CSCI 7000)

Course description and learning objectives

Resources

PhD application resources

CU Boulder's graduate admissions homepage

CU Boulder's tips for creating an impactful application and opportunities for an application fee waiver.

MOSAIKS: Generalizable and accessible machine learning with global satellite imagery

The MOSAIKS API (data interface) is now available: go to siml.berkeley.edu to make an account and use precomputed global features (using satellite imagery from 2019).

See the MOSAIKS project page for more resources and updates.