Hossein Mobahi

About Me

I am a Research Scientist in the Machine Perception team at Google as of May 2016. Prior to that, I was a Postdoctoral Researcher in Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT, where I was affiliated with the Computer Vision and the Sensing, Learning, Inference labs, and privileged to work with Profs. Bill Freeman and John Fisher.

I am broadly interested in Artificial Intelligence. Specifically my research lies at the intersection of Computer Vision, Machine Learning, and Optimization. My work is often guided by mathematical principles. I am also the admin of the facebook page on Nonconvex Optimization.

I graduated from University of Illinois at Urbana-Champaign (UIUC) with a PhD in Computer Science, where I was fortunate to be supervised by Prof. Yi Ma. My dissertation is titled Optimization by Gaussian Smoothing with Application to Geometric Alignment.

News

	New Work with Anima Anandkumar, Rong Ge, and Yuan Deng. Homotopy Analysis for Tensor PCA. Conference On Learning Theory (COLT) 2017.
	We are organizing the NIPS 2016 workshop on nonconvex optimization. [URL].
	The video of my talk at the NIPS 2015 workshop on nonconvex optimization is now online. [Youtube].
	An overview of my NIPS 2015 invited talk on nonconvex optimization. Off The Convex Optimization.
	I recently gave an invited talk at NIPS 2015. Workshop on Nonconvex Optimization for Machine Learning.
	Paper accepted in NIPS 2015. Learning with a Wasserstein Loss.
	Paper accepted in CVPR 2015. The Aperture Problem for Refractive Motion.
	My Research about nonconvex optimization is featured at MIT News .

Selected Publications

	Homotopy Analysis for Tensor PCA [Paper] [BibTeX] Anima Anandkumar, Yuan Deng, Rong Ge, Hossein Mobahi (authors ordered alphabetically). Conference On Learning Theory (COLT), 2017. We analyze homotopy continuation for global optimization of nonconvex functions. In particular, we prove global convergence of the homotopy method for tensor PCA problem in the "high noise" regime. The signal-to-noise requirement for our algorithm is tight; it matches the recovery guarantee for the best degree-4 sum-of-squares algorithm.
	Training Recurrent Neural Networks by Diffusion [Paper] [BibTeX] [arXiv] [Talk] Hossein Mobahi. arXiv:1601.04114 [cs.LG], January 2016. Diffusion theory can justify some of the common techniques used in deep learning, and also lead to much faster training compared to SGD. The arXiv version may be lagged in updates. Please use the local copy (linked as "Paper") for the most recent version.
	Learning With a Wasserstein Loss [Paper] [Supplement] [BibTeX] [Project Page] Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya-Polo, Tomaso Poggio. Neural Information Processing Systems (NIPS), 2015. Multi-label learning via Wasserstein distance with efficient computation. Improved results for tag prediction problem with Yahoo Flickr dataset.
	The Aperture Problem for Refractive Motion [Paper] [BibTeX] [Project Page] Tianfan Xue, Hossein Mobahi, Fredo Durand, William T. Freeman. Int. Conf. on Computer Vision and Pattern Recognition (CVPR), 2015. Extending the theory of the aperture problem -- fundamental in motion estimation -- to more challenging setup involving refraction.
	A Theoretical Analysis of Optimization by Gaussian Continuation [Paper] [BibTeX] Hossein Mobahi, John W. Fisher III. 29th Conference on Artificial Intelligence (AAAI), 2015. For the first time, performance guarantees are provided for optimization by continuation by combining regularization theory and differential equations.
	On the Link Between Gaussian Homotopy Continuation and Convex Envelopes [Paper] [BibTeX] [MIT News] Hossein Mobahi, John W. Fisher III. Energy Minimization Method in Computer Vision and Pattern Recognition (EMMCVPR), 2015. For the first time, an optimal homotopy construction for optimization by continuation is derived via partial differential equations.
	A Compositional Model for Low-Dimensional Image Set Representation [Paper] [BibTeX] [Video MPG(57MB)] Hossein Mobahi, Ce Liu, William T. Freeman. Int. Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. Discovering the space of variations within an image set by nonlinear decomposition of images into low-dimensional components. Results on image morphing, motion synthesis, and navigation through images are presented.
	Seeing through the Blur [Paper] [BibTeX] [Project Page] Hossein Mobahi, C. Lawrence Zitnick, Yi Ma. Int. Conf. on Computer Vision and Pattern Recognition (CVPR), 2012. Mathematical derivation of optimal blur operators for image alignment. The resulted blur is spatilly varying like in Retina, unlike the widely used uniform blur in multiresolution methods.
	Segmentation of Natural Images by Texture and Boundary Compression [Paper] [BibTeX] [Project Page] Hossein Mobahi, Shankar R. Rao, Allen Y. Yang, S. Shankar Sastry, Yi Ma. International Journal of Computer Vision (IJCV), 2011. State of the art segmentation result on standard benchmark dataset via Gaussian source coding of information theory.
	Deep Learning from Temporal Coherence in Video [Paper] [BibTeX] [Project Page] [Dataset] [Slides] [Video Lecture] Hossein Mobahi, Ronan Collobert, Jason Weston. Int. Conf. on Machine Learning (ICML), 2009. One of the first papers on deep learning from video to improve recognition. This exploits the naturally existing temporal coherence among successive video frames.
	Builidng an Interactive Robot Face from Scratch [Paper] [BibTeX] [Project Page] Hossein Mobahi. Bachelors Degree Final Project. Showing how an intelligent interactive robot face can be built from scratch. The electronic and hardware is built from basic elements and the software is coded up in c and includes various modules including neural networks and PID control. The robot can detect faces and hands, and follow them by adjusting its gaze and showing emotions via motorized mouth and eyebrows.