About
I am an Assistant Professor in the Department of Computer Science at the University of Chicago. I have a courtesy appointment at the Toyota Technological Institute at Chicago (TTIC).

Previously, I was a Research Assistant Professor at TTIC, following a time as a postdoc in the Computer Vision Lab at Caltech. I completed my Ph.D. at UC Berkeley, advised by Jitendra Malik.
Contact
The University of Chicago
Department of Computer Science
John Crerar Library Building
5730 S. Ellis Ave
Chicago, IL 60637

mmaire@uchicago.edu
Students

Information-Theoretic Segmentation by Inpainting Error Maximization

Pedro Savarese, Sunnie S. Y. Kim, Michael Maire, Greg Shakhnarovich, and David McAllester
Computer Vision and Pattern Recognition (CVPR), 2021
An optimization approach to unsupervised image segmentation that minimizes predictability of foreground from background and vice-versa, optionally producing pseudo-labels for training deep models.

Domain-Independent Dominance of Adaptive Methods

Pedro Savarese, David McAllester, Sudarshan Babu, and Michael Maire
Computer Vision and Pattern Recognition (CVPR), 2021
AvaGrad: an adaptive optimizer that matches or outperforms both Adam and SGD, simplifies hyperparameter tuning by decoupling learning rate from adaptivity, and improves training of GANs.

Multimodal Contrastive Training for Visual Representation Learning

Xin Yuan, Zhe Lin, Jason Kuen, Jianming Zhang, Yilin Wang, Michael Maire, Ajinkya Kale, and Baldo Faieta
Computer Vision and Pattern Recognition (CVPR), 2021
A self-supervised learning framework connecting vision and language in a common embedding space, with both inter-modal and intra-modal contrastive objectives.

Are Machine Learning Cloud APIs Used Correctly?

Chengcheng Wan, Shicheng Liu, Henry Hoffmann, Michael Maire, and Shan Lu
International Conference on Software Engineering (ICSE), 2021
An analysis of open-source applications that use cloud-based machine learning APIs, in combination with development of automated checkers for identifying API misuses.

Growing Efficient Deep Networks by Structured Continuous Sparsification

Xin Yuan, Pedro Savarese, and Michael Maire
International Conference on Learning Representations (ICLR), 2021 [Oral]
Accelerated training of compact neural networks by growing them from a small seed architecture, using a combination of continuous sparsification with a sampling scheme; unifies pruning and architecture search.

Winning the Lottery with Continuous Sparsification

Pedro Savarese*, Hugo Silva*, and Michael Maire
Neural Information Processing Systems (NeurIPS), 2020
A singular approach to pruning neural networks and finding lottery tickets (sparse subnetworks, retrainable from initialization), with experimental results setting a new state-of-the-art for each.

Self-Supervised Visual Representation Learning from Hierarchical Grouping

Xiao Zhang and Michael Maire
Neural Information Processing Systems (NeurIPS), 2020 [Spotlight]
A generic image segmentation procedure, trained on small dataset, bootstraps self-supervised representation learning on a large unlabeled dataset using a region-aware contrastive loss.

Multigrid Neural Memory

Tri Huynh, Michael Maire, and Matthew R. Walter
International Conference on Machine Learning (ICML), 2020
A design for endowing neural networks with hierarchically-connected distributed memory; end-to-end training yields the emergence of coherent memory subsystems driven by learned addressing strategies.

Orthogonalized SGD and Nested Architectures for Anytime Neural Networks

Chengcheng Wan, Henry Hoffmann, Shan Lu, and Michael Maire
International Conference on Machine Learning (ICML), 2020
The principle of maximizing re-use of intermediate representations within a multitask network guides design of multistage nested network architectures paired with a custom training procedure.

ALERT: Accurate Anytime Learning for Energy and Timeliness

Chengcheng Wan, Muhammad Santriaji, Eri Rogers, Henry Hoffmann, Michael Maire, and Shan Lu
USENIX Annual Technical Conference (USENIX ATC), 2020
A runtime scheduler for selecting neural network and system resource configurations to meet latency, accuracy, and energy constraints in dynamic operating environments.

Pixel Consensus Voting for Panoptic Segmentation

Haochen Wang, Ruotian Luo, Michael Maire, and Greg Shakhnarovich
Computer Vision and Pattern Recognition (CVPR), 2020
Object segmentation based on the generalized Hough transform, which accumulates pixelwise votes for instance centroids over discrete spatial cells, implemented within a deep convolutional framework.

Learning Implicitly Recurrent CNNs Through Parameter Sharing

Pedro Savarese and Michael Maire
International Conference on Learning Representations (ICLR), 2019
Re-parameterizing convolutional networks, so that layers share weights via a set of templates, benefits generalization and often produces trained models that can be rewritten in an explicitly recurrent form.

Sparsely Aggregated Convolutional Networks

Ligeng Zhu, Ruizhi Deng, Michael Maire, Zhiwei Deng, Greg Mori, and Ping Tan
European Conference on Computer Vision (ECCV), 2018
Scaling skip or residual connection scaffolding logarithmically, rather than linearly (standard ResNet), with network depth increases parameter efficiency and permits training models thousands of layers deep.

Self-Supervised Relative Depth Learning for Urban Scene Understanding

Huaizu Jiang, Erik Learned-Miller, Gustav Larsson, Michael Maire, and Greg Shakhnarovich
European Conference on Computer Vision (ECCV), 2018
Predicting relative depth from a single source image serves as proxy task for self-supervised learning; trained from egomotion video, these learned representations boost several downstream tasks.

Regularizing Deep Networks by Modeling and Predicting Label Structure

Mohammadreza Mostajabi, Michael Maire, and Gregory Shakhnarovich
Computer Vision and Pattern Recognition (CVPR), 2018
Learning an autoencoder for a structured label space (e.g., semantic segmentation annotations), allows subsequent use of its decoder as an auxiliary output branch to regularize training of a prediction network.

Colorization as a Proxy Task for Visual Understanding

Gustav Larsson, Michael Maire, and Gregory Shakhnarovich
Computer Vision and Pattern Recognition (CVPR), 2017
Requiring no labeled data, colorizing images serves as an effective training proxy for downstream classification and semantic segmentation tasks; we quantify in comparison to ImageNet pretraining.

Multigrid Neural Architectures

Tsung-Wei Ke, Michael Maire, and Stella X. Yu
Computer Vision and Pattern Recognition (CVPR), 2017
Endowing CNNs with spatial shortcuts, by remaking layers into multigrid pyramids with bidirectional wiring across grid scales, yields an emergent capability: attentional behavior arises from end-to-end training.

FractalNet: Ultra-Deep Neural Networks without Residuals

Gustav Larsson, Michael Maire, and Gregory Shakhnarovich
International Conference on Learning Representations (ICLR), 2017
A self-similar architecture shows that shortcut pathways across depth, not the functional form of residual connections, are key to training networks, allowing them to transition from effectively shallow to deep.

Learning Representations for Automatic Colorization

Gustav Larsson, Michael Maire, and Gregory Shakhnarovich
European Conference on Computer Vision (ECCV), 2016 [Oral]
A hypercolumn architecture and a loss formulated on histogram prediction serve as the basis for training a colorization network that outperforms existing methods and yields broadly useful feature representations.

Affinity CNN: Learning Pixel-Centric Pairwise Relations for Figure/Ground Embedding

Michael Maire, Takuya Narihira, and Stella X. Yu
Computer Vision and Pattern Recognition (CVPR), 2016 [Spotlight]
Spectral embedding resolves predictions, produced by a CNN, about local figure/ground and grouping relationships into a globally consistent segmentation and figure/ground organization of a scene.

Direct Intrinsics: Learning Albedo-Shading Decomposition by Convolutional Regression

Takuya Narihira, Michael Maire, and Stella X. Yu
International Conference on Computer Vision (ICCV), 2015
Trained on the synthetic MPI Sintel dataset to directly predict albedo and shading channels from an input RGB image, a convolutional neural network generalizes well to producing decompositions of real images.

Learning Lightness from Human Judgement on Relative Reflectance

Takuya Narihira, Michael Maire, and Stella X. Yu
Computer Vision and Pattern Recognition (CVPR), 2015
An approach to predicting relative lightness (perceived reflectance) of image patches using deep features, with promising results on the Intrinsic Images in the Wild dataset.

Reconstructive Sparse Code Transfer for Contour Detection and Semantic Labeling

Michael Maire, Stella X. Yu, and Pietro Perona
Asian Conference on Computer Vision (ACCV), 2014 [Oral]
Deep, multiscale feature representations based on sparse coding; articulates the concept of concatenating a spatially-localized slice through multiple layers of sparse codes (activations) into a feature descriptor.

Microsoft COCO: Common Objects in Context

Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick
European Conference on Computer Vision (ECCV), 2014 [Oral]
A large-scale dataset for object detection and segmentation.

Progressive Multigrid Eigensolvers for Multiscale Spectral Segmentation

Michael Maire and Stella X. Yu
International Conference on Computer Vision (ICCV), 2013
A custom eigensolver, using a novel interpolation strategy, for spectral clustering problems with coarse-to-fine structure and (optional) constraints; applied to image segmentation, it delivers a substantial speedup.

Hierarchical Scene Annotation

Michael Maire, Stella X. Yu, and Pietro Perona
British Machine Vision Conference (BMVC), 2013
Web-based software tools for detailed annotation of scenes into object-part-subpart region hierarchies and their occlusion relationships.

Object Detection and Segmentation from Joint Embedding of Parts and Pixels

Michael Maire, Stella X. Yu, and Pietro Perona
International Conference on Computer Vision (ICCV), 2011 [Oral]
A model of pixel grouping, part grouping, and figure/ground relationships as interactions between graph nodes, with a globally consistent scene interpretation obtained by solving an Angular Embedding problem.

Occlusion Boundary Detection and Figure/Ground Assignment from Optical Flow

Patrik Sundberg, Thomas Brox, Michael Maire, Pablo Arbeláez, and Jitendra Malik
Computer Vision and Pattern Recognition (CVPR), 2011 [Oral]
A contour detection procedure for video that exploits motion cues and, in combination with reasoning about optical flow over a segmentation into regions, determines figure/ground relationships.

Contour Detection and Hierarchical Image Segmentation

Pablo Arbeláez, Michael Maire, Charless Fowlkes, and Jitendra Malik
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2011
A high-performance contour detector, combining local and global information, and generic machinery for transforming contours into a hierarchical region tree; extensive quantitative evaluation.

Simultaneous Segmentation and Figure/Ground Organization using Angular Embedding

Michael Maire
European Conference on Computer Vision (ECCV), 2010
Angular Embedding, an extension of spectral clustering to complex-valued affinities, provides the basis for incorporating figure/ground ordering relationships into an image segmentation framework.

From Contours to Regions: An Empirical Evaluation

Pablo Arbeláez, Michael Maire, Charless Fowlkes, and Jitendra Malik
Computer Vision and Pattern Recognition (CVPR), 2009
An Oriented variant of the Watershed Transform (OWT), followed by construction of an Ultrametic Contour Map (UCM), together serve to convert contours into a region hierarchy, while preserving boundary quality.

Using Contours to Detect and Localize Junctions in Natural Images

Michael Maire, Pablo Arbeláez, Charless Fowlkes, and Jitendra Malik
Computer Vision and Pattern Recognition (CVPR), 2008 [Oral]
A method for enhancing local contour detection with global information obtained from spectral partitioning, and a novel iterative approach to localizing junctions from contour fragments.

SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition

Hao Zhang, Alexander C. Berg, Michael Maire, and Jitendra Malik
Computer Vision and Pattern Recognition (CVPR), 2006
A hybridization of support vector machines (SVMs) and nearest neighbor classification, achieved by training a local SVM on the collection of neighbors close to a query point.

Making Latin Manuscripts Searchable using gHMM's

Jaety Edwards, Yee Whye Teh, David A. Forsyth, Roger Bock, Michael Maire, and Grace Vesom
Neural Information Processing Systems (NeurIPS), 2004
A system for making scanned, handwritten latin manuscripts accessible to full text search, while requiring only a small amount of annotated training data.

Names and Faces in the News

Tamara L. Berg, Alexander C. Berg, Jaety Edwards, Michael Maire, Ryan White, Yee Whye Teh, Erik Learned-Miller, and David A. Forsyth
Computer Vision and Pattern Recognition (CVPR), 2004
A large dataset of faces obtained from captioned news images.

Recognition by Probabilistic Hypothesis Construction

Pierre Moreels, Michael Maire, and Pietro Perona
European Conference on Computer Vision (ECCV), 2004
A probabilistic framework for recognizing individual object instances in images of cluttered scenes.
Theses

Contour Detection and Image Segmentation

Michael Maire
Ph.D. Thesis
University of California, Berkeley, 2009

Dynamic Code Updates

Michael Maire
Undergraduate Thesis
California Institute of Technology, 2003