We study the supervised learning problem under either of the following two models: (1) Feature vectors xi are ddimensional Gaussian and responses are yi = f*(xi) for f* an unknown quadratic function; (2) Feature vectors xi are distributed as a mixture of two ddimensional centered Gaussians, and y_i's are the corresponding class labels. We use twolayers neural networks with quadratic activations, and compare three different learning regimes: the random features (RF) regime in which we only train the secondlayer weights; the neural tangent (NT) regime in which we train a linearization of the neural network around its initialization; the fully trained neural network (NN) regime in which we train all the weights in the network. We prove that, even for the simple quadratic model of point (1), there is a potentially unbounded gap between the prediction risk achieved in these three training regimes, when the number of neurons is smaller than the ambient dimension. When the number of neurons is larger than the number of dimensions, the problem is significantly easier and both NT and NN learning achieve zero risk.
Author Information
Song Mei (Stanford University)
Theodor Misiakiewicz (Stanford University)
Behrooz Ghorbani (Stanford University)
Andrea Montanari (Stanford)
Related Events (a corresponding poster, oral, or spotlight)

2019 Poster: Limitations of Lazy Training of Twolayers Neural Network »
Wed Dec 11th 05:00  07:00 PM Room East Exhibition Hall B + C
More from the Same Authors

2018 Poster: Contextual Stochastic Block Models »
Yash Deshpande · Subhabrata Sen · Andrea Montanari · Elchanan Mossel 
2018 Spotlight: Contextual Stochastic Block Models »
Yash Deshpande · Subhabrata Sen · Andrea Montanari · Elchanan Mossel 
2017 Poster: Inference in Graphical Models via Semidefinite Programming Hierarchies »
Murat Erdogdu · Yash Deshpande · Andrea Montanari 
2015 Poster: Convergence rates of subsampled Newton methods »
Murat Erdogdu · Andrea Montanari 
2015 Poster: On the Limitation of Spectral Methods: From the Gaussian Hidden Clique Problem to RankOne Perturbations of Gaussian Tensors »
Andrea Montanari · Daniel Reichman · Ofer Zeitouni 
2014 Poster: A statistical model for tensor PCA »
Emile Richard · Andrea Montanari 
2014 Poster: ConeConstrained Principal Component Analysis »
Yash Deshpande · Andrea Montanari · Emile Richard 
2014 Poster: Sparse PCA via Covariance Thresholding »
Yash Deshpande · Andrea Montanari 
2013 Poster: Estimating LASSO Risk and Noise Level »
Mohsen Bayati · Murat Erdogdu · Andrea Montanari 
2013 Poster: Confidence Intervals and Hypothesis Testing for HighDimensional Statistical Models »
Adel Javanmard · Andrea Montanari 
2013 Poster: Model Selection for HighDimensional Regression under the Generalized Irrepresentability Condition »
Adel Javanmard · Andrea Montanari 
2010 Poster: Learning Networks of Stochastic Differential Equations »
José Bento · Morteza Ibrahimi · Andrea Montanari 
2010 Poster: The LASSO risk: asymptotic results and real world examples »
Mohsen Bayati · José Bento · Andrea Montanari 
2009 Poster: Matrix Completion from Noisy Entries »
Raghunandan Keshavan · Andrea Montanari · Sewoong Oh 
2009 Poster: Which graphical models are difficult to learn? »
Andrea Montanari · José Bento