Towards Provable Learning of Polynomial Neural Networks Using Low-Rank Matrix Estimation

Mohammadreza Soltani, Chinmay Hegde
Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics, PMLR 84:1417-1426, 2018.

Abstract

We study the problem of (provably) learning the weights of a two-layer neural network with quadratic activations. In particular, we focus on the under-parametrized regime where the number of neurons in the hidden layer is (much) smaller than the dimension of the input. Our approach uses a lifting trick, which enables us to borrow algorithmic ideas from low-rank matrix estimation. In this context, we propose two novel, non-convex training algorithms which do not need any extra tuning parameters other than the number of hidden neurons. We support our algorithms with rigorous theoretical analysis, and show that the proposed algorithms enjoy linear convergence, fast running time per iteration, and near-optimal sample complexity. Finally, we complement our theoretical results with several numerical experiments.

Cite this Paper


BibTeX
@InProceedings{pmlr-v84-soltani18a, title = {Towards Provable Learning of Polynomial Neural Networks Using Low-Rank Matrix Estimation}, author = {Soltani, Mohammadreza and Hegde, Chinmay}, booktitle = {Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics}, pages = {1417--1426}, year = {2018}, editor = {Storkey, Amos and Perez-Cruz, Fernando}, volume = {84}, series = {Proceedings of Machine Learning Research}, month = {09--11 Apr}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v84/soltani18a/soltani18a.pdf}, url = {https://proceedings.mlr.press/v84/soltani18a.html}, abstract = {We study the problem of (provably) learning the weights of a two-layer neural network with quadratic activations. In particular, we focus on the under-parametrized regime where the number of neurons in the hidden layer is (much) smaller than the dimension of the input. Our approach uses a lifting trick, which enables us to borrow algorithmic ideas from low-rank matrix estimation. In this context, we propose two novel, non-convex training algorithms which do not need any extra tuning parameters other than the number of hidden neurons. We support our algorithms with rigorous theoretical analysis, and show that the proposed algorithms enjoy linear convergence, fast running time per iteration, and near-optimal sample complexity. Finally, we complement our theoretical results with several numerical experiments.} }
Endnote
%0 Conference Paper %T Towards Provable Learning of Polynomial Neural Networks Using Low-Rank Matrix Estimation %A Mohammadreza Soltani %A Chinmay Hegde %B Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2018 %E Amos Storkey %E Fernando Perez-Cruz %F pmlr-v84-soltani18a %I PMLR %P 1417--1426 %U https://proceedings.mlr.press/v84/soltani18a.html %V 84 %X We study the problem of (provably) learning the weights of a two-layer neural network with quadratic activations. In particular, we focus on the under-parametrized regime where the number of neurons in the hidden layer is (much) smaller than the dimension of the input. Our approach uses a lifting trick, which enables us to borrow algorithmic ideas from low-rank matrix estimation. In this context, we propose two novel, non-convex training algorithms which do not need any extra tuning parameters other than the number of hidden neurons. We support our algorithms with rigorous theoretical analysis, and show that the proposed algorithms enjoy linear convergence, fast running time per iteration, and near-optimal sample complexity. Finally, we complement our theoretical results with several numerical experiments.
APA
Soltani, M. & Hegde, C.. (2018). Towards Provable Learning of Polynomial Neural Networks Using Low-Rank Matrix Estimation. Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 84:1417-1426 Available from https://proceedings.mlr.press/v84/soltani18a.html.

Related Material