Colorado Reed

echo 'bW9jLmxpYW1nQDczMzErZGVlcmpvZGFyb2xvYwo=' | base64 --decode | rev

I'm Colorado Reed (colorado _dot_ j _dot_ reed _at_ gmail): an Applied Research Scientist at Apple. I completed my Ph.D. at Berkeley in the Berkeley Artificial Intelligence Research Lab and was also a researcher at Meta AI.

I was previously a cofounder and CTO at a venture-backed fintech startup for 5 years, where I led an engineering team in building and deploying products for Fortune 500 clients. I also co-founded and led the Berkeley AI Research Climate Initiative. I am a champion of democratized education. Along with Roger Grosse, I spent years building an open-source system that creates personalized learning plans for students focused on machine learning and statistics concepts (see metacademy.org).

→ t e l l m e m o r e . . . ←

Your name is really "Colorado"?

Yup. Since birth. Sharing a name with a state is surprisingly common, see Tennessee Williams, Indiana Jones, Virginia Woolf, Dakota Fanning, Georgia O'Keeffe, Tex Ritter, or York Bowen ;-).

Why do you use "Colorado J Reed"?

Google doesn't think Colorado Reed is a person, so it removes that name from the autogenerated bibtex entries for my publications. Adding my middle initial seems to have solved this problem.

Is it true that Google is successful because of you?

True story: As a kid (around 1999/2000), I was watching a show on PBS where they mentioned the number 'googol' (1e100) -- they showed some guy writing googol on a chalkboard. Later that day, I went to the library to use their internet (we didn't have a computer). and I decided to check out googol.com, which returned an error, so I thought I mispelled googol and tried googel, googal, and google. The last one was, as you probably know, a search engine. It returned really good results and was much cleaner than AltaVista and Yahoo, so I started using it regularly. I also told everyone I knew about this great new search engine I stumbled across. Over the following years I watched google rise to prominence, and I was convinced this was at least partially my doing (word-of-mouth and all).

Publications

Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning

Colorado J Reed*, Ritwik Gupta*, Shufan Li*, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uyttendaele, Trevor Darrell

ICCV 2023. [preprint].

Learn representations for geospatial data by explicitly conditioning on its scale and autoencoding the data to different resolutions.
Refine and Represent: Region-to-Object Representation Learning

Akash Gokul*, Konstantinos Kallidromitis*, Shufan Li*, Yusuke Kato, Kazuki Kozuka, Trevor Darrell, Colorado J Reed

[preprint]. 2022.

Representation learning benefits by first learning from image regions and then learning from actual objects.
Prior Knowledge-Guided Attention in Self-Supervised Vision Transformers

Kevin Miao, Akash Gokul, Raghav Singh, Suzanne Petryk, Joseph Gonzalez, Kurt Keutzer, Trevor Darrell, Colorado J Reed

[preprint]. 2022.

We know all sorts of spatial relationships for certain kinds of data. For example, with medical images, lungs tend to be next to the heart and above the stomach. We present a mechanism to incorporate this type of known spatial relationships during representation learning.
DETReg: Unsupervised Pretraining with Region Priors for Object Detection.

Amir Bar, Xin Wang, Vadim Kantorov, Colorado J Reed, Roei Herzig, Gal Chechik, Anna Rohrbach, Trevor Darrell, Amir Globerson

Conference on Computer Vision and Pattern Recognition (CVPR), 2022.

[paper]. [project page]

Using classic CV techniques (selective search) provides a useful form of "supervision" for unsupervised object detection.
Self-Supervised Pretraining Improves Self-Supervised Pretraining

Colorado J Reed*, Xiangyu Yue*, Ani Nrusimha, Sayna Ebrahimi, Vivek Vijaykumar, Richard Mao, Bo Li, Shanghang Zhang, Devin Guillory, Sean Metzger, Kurt Keutzer, Trevor Darrell

WACV 2022.

[paper] [code/models]

Initializing self-supervised pretraining with a pretrained model is pretty much always a good idea.
Invariant Information Bottleneck for Domain Generalization

Bo Li, Yifei Shen, Yezhen Wang, Wenzhen Zhu, Colorado J Reed, Jun Zhang, Dongsheng Li, Kurt Keutzer, Han Zhao

AAAI 2022.

[paper]

This paper proposes a new objective function that works towards domain generalization by minimizing the mutual information between input data and learned representations.
Region Similarity Representation Learning.

Tete Xiao*, Colorado J Reed*, Xiaolong Wang, Kurt Keutzer, Trevor Darrell

International Conference on Computer Vision (ICCV), 2021.

[paper] [code/models]

Maintain spatial relationships in the convolutional feature maps when performing instance contrastive pretraining -- this is useful for detection-related tasks.
SelfAugment: Automatic Augmentation Policies for Self-Supervised Learning.

Colorado J Reed*, Sean Metzger*, Aravind Srinivas, Trevor Darrell, Kurt Keutzer

Conference on Computer Vision and Pattern Recognition (CVPR), 2021.

[paper] [code]

Determine strong augmentation policies for instance contrastive learning without using supervised evaluations.
Video Digests: A Browsable, Skimmable Format for Informational Lecture Videos.

Amy Pavel, Colorado J Reed, Bjoern Hartmann, and Maneesh Agrawala

ACM 27th Symposium on User Interface Software and Technology, 2014.

[pdf] [code] [website] [demo]

Browse an informational video just like a textbook.
Scaling the Indian Buffet Process via Submodular Maximization.

Colorado J Reed, Zoubin Ghahramani

International Conference on Machine Learning (ICML), 2013.

[pdf] [code]

Use submodular optimization to shortcut training of latent feature models.
Semi-supervised eigenbasis novelty detection.

David R. Thompson, Walid A. Majid, Colorado J Reed, and Kiri L. Wagstaff

Statistical Analysis and Data Mining, 2012

[pdf].

PCA is useful of novelty detection even with temporal astronomical data (expanded version of below workshop paper).
Semi-supervised novelty detection with adaptive eigenbases, and application to radio transients.

David R. Thompson, Walid A. Majid, Colorado J Reed, and Kiri L. Wagstaff

Conference on Intelligent Data Understanding, 2011. Best Paper Award

[pdf]

PCA is useful for novelty detection even with temporal astronomical data.

Workshops

Snowpack Estimation in Key Mountainous Water Basins from Openly-Available, Multimodal Data Sources

Malachy Moran, Kayla Woputz, Derrick Hee, Manuela Girotto, Paolo D'Odorico, Ritwik Gupta, Daniel Feldman, Puya Vahabi, Alberto Todeschini, Colorado J Reed

CVPR MultiEarth Workshop - Oral Presentation, 2022.

[pdf]

Using multiple streams of geospatial data to estimate snowpack in the Sierra Nevadas.
SunCast: Solar Irradiance Nowcasting from Geosynchronous Satellite Data

Dhileeban Kumaresan, Richard Wang, Ernesto A Martinez, Richard Cziva, Alberto Todeschini, Colorado J Reed, Puya Vahabi

NeurIPS 2021 Workshop on Tackling Climate Change with Machine Learning.

[paper]

Improving solar nowcasting in multiple geographical regions.
HyperionSolarNet: Solar Panel Detection from Aerial Images

Poonam Parhar, Ryan Sawasaki, Alberto Todeschini, Colorado J Reed, Hossein Vahabi, Nathan Nusaputra, Felipe Vergara

NeurIPS 2021 Workshop on Tackling Climate Change with Machine Learning.

[paper]

Detecting solar panels and building an associated dataset from aerial images.
Self-supervised Contrastive Learning for Irrigation Detection in Satellite Imagery

Chitra Agastya, Sirak Ghebremusse, Ian Anderson, Colorado J Reed, Hossein Vahabi, Alberto Todeschini

ICML Workshop Tackling Climate Change with Machine Learning, 2021.

[paper]

Self-supervised pretraining works relatively well for irrigation detection in satellite imagery + we did a real-world case study.
What's trending? Mining topical trends in UGC systems with YouTube as a case study.

Colorado J Reed, Todd Elvers, and Padmini Srinivasan

17th ACM SIGKDD, 2011.

[pdf]

Create clusters of emerging topics on Twitter, YouTube, etc.

Tech Reports

Multi-source Few-shot Domain Adaptation

Xiangyu Yue, Zangwei Zheng, Colorado J Reed, Hari Prasanna Das, Kurt Keutzer, Alberto Sangiovanni Vincentelli

tech report, 2021.

A domain adaptation method for learning from multiple sources of labeled data when there are a few labeled examples in the target domain.
Multi-source Domain Adaptation in the Deep Learning Era: A Systematic Survey.

Sicheng Zhao, Bo Li, Colorado J Reed, Pengfei Xu, Kurt Keutzer

tech report, 2020.

A recent survey on multi-source domain adaptation.
Submodular MAP Inference for Scalable Latent Feature Models.

Colorado J Reed

Master's Thesis, University of Cambridge, 2013.

[pdf] [code]

Use submodular optimization to shortcut training, described in 100+ glorious pages.
Latent Dirichlet Allocation: Towards a Deeper Understanding.

Colorado J Reed. 2012.

[pdf] [code]

A tutorial on why LDA works.

Advising and Mentoring

I've had the pleasure of advising and mentoring a number of students directly:

Akash Gokul (2021-2022; BAIR Masters; Now at Salesforce AI Research)
Araav Patel (2021-Present; BAIR Undergraduate)
Jacob Yeung (2022-Present; BAIR Undergraduate)
Kevin Miao (2021-2022; BAIR Masters; Now at Apple)
Raymond Mo (2022-Present; BAIR Undergraduate)
Richard Mao (2020-2021; BAIR Undergraduate; Now at Meta)
Shashwath Senthil (Summer 2022; HS Student)
Shufan Li (2022-Present; BAIR Undergraduate)
Vivek Vijaykumar (2020-2021; HS Student; Now at GA Tech)

I've also served as a GSI several times (both officially and unofficially) for DATASCI W210 at UC Berkeley: a semester-long capstone project course. Through this project, I've helped guide dozens of project groups, leading to many of the workshop publications above =).

Some Awards I've Been Fortunate to Receive

Winston Churchill Scholarship
NDSEG Fellowship
NSF Graduate Fellowship ([politely] declined for NDSEG)
U. Iowa Ernest R. Johnson Memorial Prize (valedictorian at University of Iowa)
Barry M. Goldwater Scholarship