Lecture 7
Apart from their intrinsic interest, moment/tensor methods are perfectly parallelizable. Moments can be computed in a single pass with no communication.
Readings:
Tensor Decompositions for Learning Latent Variable Models Links to an external site. By A. Anandkumar, R. Ge, D. Hsu, S.M. Kakade and M. Telgarsky. Journal of Machine Learning Research 15 (2014) 2773-2832
This is a long paper, so skip the appendices and proofs
Slides Download Slides by Carlos Florensa
Beating the Perils of Non-Convexity: Guaranteed Training of Neural Networks using Tensor Methods Links to an external site., by M. Janzamin, H. Sedghi and A. Anandkumar, June. 2015
Slides Download Slides by Yuansi Chen
Provable Methods for Training Neural Networks with Sparse Connectivity Links to an external site., by H. Sedghi and A. Anandkumar, Neural Information Processing Systems (NIPS) Deep Learning Workshop, 2014 and in International Conference on Learning Representation (ICLR), May, 2015
Slides Download Slides by Rohin Shah
Recommended:
Reinforcement Learning of POMDPs using Spectral Methods Links to an external site. By Kamyar Azizzadenesheli, Alessandro Lazaric, Anima Anandkumar, 2016
Global Optimality in Tensor Factorization Links to an external site., Deep Learning, and Beyond, Benjamin D. Haeffele, Rene Vidal, arXiv 1506.07540v1, 2015.