Michael G. Ross

mgross@mit.edu (617) 258-9675 46-4078B Brain and Cognitive Sciences Massachusetts Institute of Technology Curriculum vitae

I am a computer scientist and my primary research interest is the application of statistical machine learning techniques to challenging problems in computer vision, human vision, and other real-world domains.

Currently, I am a postdoctoral associate at the MIT Department of Brain and Cognitive Sciences. I am a member of Aude Oliva's Computational Visual Cognition Laboratory

Previously, I have worked with:

• Andrew Cohen (at the UMass-Amherst Psychology Department)
• Leslie Pack Kaelbling (at the MIT Computer Science and Artificial Intelligence Laboratory) [Ph.D. advisor]
• Paul Viola (at the MIT Aritifical Intelligence Laboratory) [S.M. advisor]
• Daniela Rus (at the Dartmouth College Computer Science Department) [A.B. advisor]

Current Research

• Human perception and visual feature spaces (2007-present)
  Researchers in computer vision and cognitive psychology have developed many features for image classification and recognition algorithms. This project explores the relationships between some of these feature spaces and human perception. Collaborators: Emmanuelle Boloix, Aude Oliva
• Inducing visual classification features (2005-present)
  What features do humans use to classify images? Much previous work on this question assumes that people are applying a linear classifier to image pixels. Our new approach assumes a two-stage classification process - probabilistic feature detection followed by classification - and uses machine learning algorithms for graphical model parameter learning to find the features. We are also exploring the use of more sophisticated features and the recognition of natural scenes.
  Collaborators: Andrew Cohen, Michelle Greene
  Publications:
  • Michael G. Ross, Andrew L. Cohen. GRIFT: A graphical model for inferring visual classification features from human data. Advances in Neural Information Processing Systems (NIPS) 20. 2008.
  • Andrew L. Cohen, Richard M. Shiffrin, Jason M. Gold, David A. Ross, Michael G. Ross. Inducing features from visual noise. Journal of Vision, 7(8). 2007.
  Presentations:
  • Michael Ross and Andrew Cohen. Modeling human visual classification with Bayesian networks. Sixth Annual Summer Interdisciplinary Conference (ASIC). 2007.
  • Michael Ross and Andrew Cohen. Inducing multiple independent classification templates from stochastic stimuli. Fifth Annual Summer Interdisciplinary Conference (ASIC). 2006.
• Markov chain Monte Carlo analysis of human mass perception (2006-present)
  Using a technique developed by Tom Griffiths and Adam Sanborn, we are using Markov chain Monte Carlo sampling to better understand the human perception of mass in collisions.
  Collaborators: Andrew Cohen
  Presentations:
  • Andrew Cohen and Michael Ross. Exploring Mass Ratio Perception with Psychological MCMC. Sixth Annual Summer Interdisciplinary Conference (ASIC). 2007.

Previous Research

• Iterative contextual modeling for optical character recognition (2006-2007)
  This is a new approach to optical character recognition (OCR) based on building dynamic document-specific models. Previous OCR algorithms have relied on global static document, language, and font models that are extremely expensive to construct and fail to adapt to specialized vocabulary, atypical fonts, or unusual formatting. Many of these problems can be solved by dynamically constructing a model during the recognition process that accounts for vocabulary, font, and formatting consistencies within a single document or corpus of related documents.
  Collaborators: Erik Learned-Miller, Michael Wick, Jerod Weinman
  Publications:
  • Michael Wick, Michael G. Ross, Erik Learned-Miller. Context-Sensitive Error Correction: Using Topic Models to Improve OCR. Ninth International Conference on Document Analysis and Recognition (ICDAR). September 2007.
• Event-related potential signal alignment and modeling (2006-2007)
  Cognitive psychologists study the EEG signals produced by different stimuli and tasks to understand the underlying cognitive processes. These event-related potentials (ERPs) are contaminated with noise from unrelated brain or muscle functions and from the recording and measurement process. The most common method of reducing noise is averaging large numbers of trials, but this can cause significant loss of data, particularly in cases where the latency of ERP components is highly variable. We are attempting to use machine learning and information theory techniques to better account for latency variation.
  Collaborators: Marwan Mattar, Lisa Sanders, Erik Learned-Miller
• Using eye movements to understand complex visual comparisons (2005-2006)
  This project employs the statistical analysis of eye-movement patterns to understand how humans determine the level of similarity between two graphs.
  Publications:
  • Kyle R. Cave, Andrew Cohen, Caren Rotello, Anthony McCaffrey, Michael G. Ross, Min Zeng, Xingshan Li, Matthew Zivot, Kris Chang. Using Eye Movements to Understand Complex Visual Comparisons. (In press)
• Learning Object Boundary Detection from Object Motion (2000-2005) [MIT Ph.D. Thesis]
  Advisor:Leslie Pack Kaelbling
  Publications:
  • Michael G. Ross, Leslie Pack Kaelbling. Segmentation According to Natural Examples: Learning Static Segmentation from Motion Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence. (In press)
  • Michael G. Ross. Learning Static Object Segmentation from Motion Segmentation. Massachusetts Institute of Technology Ph.D. Thesis. August 2005.
  • Michael G. Ross, Leslie Pack Kaelbling. Learning Static Object Segmentation from Motion Segmentation. Twentieth National Conference on Artificial Intelligence. July 2005.
  • Michael G. Ross, Leslie Pack Kaelbling. A Systematic Approach to Learning Object Segmentation from Motion. NIPS 2003 Workshop on Open Challenges in Cognitive Vision. December 2003.
  • Michael G. Ross, Leslie Pack Kaelbling. Learning object segmentation from video data. MIT Artificial Intelligence Lab Memo, AIM-2003-022 (MIT-CSAIL-TR-2003-018). September 2003.
  Presentations:
  • Learning object segmentation from motion segmentation December 2004
  • A Systematic Approach to Learning Object Segmentation from Motion December 2003
• Combining Optical Flow and Image Segmentation (1998-2000) [MIT Master's Thesis]
  Advisor: Paul Viola
  Publications:
  • Michael G. Ross. Exploiting Texture-Motion Duality in Optical Flow and Image Segmentation. Massachusetts Institute of Technology Master's Thesis. August 2000.
  Code:
  • Master's Thesis Code Requires quicksort function to be implemented - see README files for details
• Distributed Robot Sensing and Manipulation (1997-1998) [Dartmouth College Senior Thesis]
  Advisor: Daniela Rus
  Publications:
  • Michael G. Ross. Multiscouting: Guiding distributed manipulation with multiple mobile sensors. Dartmouth College Technical Report PCS-TR98-332. June 1998.