Open AccessProceedings Article
Deep Neural Networks as Gaussian Processes
Jaehoon Lee,Yasaman Bahri,Roman Novak,Samuel S. Schoenholz,Jeffrey Pennington,Jascha Sohl-Dickstein +5 more
TLDR
The exact equivalence between infinitely wide deep networks and GPs is derived and it is found that test performance increases as finite-width trained networks are made wider and more similar to a GP, and thus that GP predictions typically outperform those of finite- width networks.Abstract:
It has long been known that a single-layer fully-connected neural network with an i.i.d. prior over its parameters is equivalent to a Gaussian process (GP), in the limit of infinite network width. This correspondence enables exact Bayesian inference for infinite width neural networks on regression tasks by means of evaluating the corresponding GP. Recently, kernel functions which mimic multi-layer random neural networks have been developed, but only outside of a Bayesian framework. As such, previous work has not identified that these kernels can be used as covariance functions for GPs and allow fully Bayesian prediction with a deep neural network.
In this work, we derive the exact equivalence between infinitely wide deep networks and GPs. We further develop a computationally efficient pipeline to compute the covariance function for these GPs. We then use the resulting GPs to perform Bayesian inference for wide deep neural networks on MNIST and CIFAR-10. We observe that trained neural network accuracy approaches that of the corresponding GP with increasing layer width, and that the GP uncertainty is strongly correlated with trained network prediction error. We further find that test performance increases as finite-width trained networks are made wider and more similar to a GP, and thus that GP predictions typically outperform those of finite-width networks. Finally we connect the performance of these GPs to the recent theory of signal propagation in random neural networks.read more
Citations
More filters
Posted Content
Large-scale quantum hybrid solution for linear systems of equations
M. R. Perelshtein,A. I. Pakhomchik,A. A. Melnikov,A. A. Novikov,Andreas Glatz,Gheorghe Sorin Paraoanu,V. M. Vinokur,Gordey B. Lesovik +7 more
TL;DR: In this article, a hybrid quantum algorithm for solving linear systems of equations with exponential speedup, utilizing quantum phase estimation, was proposed and implemented for solving large scale problems arising in data analysis and numerical modeling.
Isonetry : geometry of critical initializations and training
TL;DR: This work complements the existing understanding of critical initializations and shows that the curvature is proportional to the maximum singular value of the Jacobian and that optimization under orthogonality constraints ameliorates the dependence on choice of initial parameters, but is not strictly necessary.
Posted Content
Initialization of ReLUs for Dynamical Isometry
Rebekka Burkholz,Alina Dubatovka +1 more
TL;DR: In this article, the authors derive the joint signal output distribution exactly, without mean field assumptions, for fully-connected networks with Gaussian weights and biases, and analyze deviations from the mean field results.
Journal ArticleDOI
Free Dynamics of Feature Learning Processes
TL;DR: In this article , the authors present a statistical physics model of the learning process and derive an autonomous dynamical system in terms of elementary degrees of freedom of the problem determining the evolution of the relative alignment between the population matrix and the signal.
Controlling Neural Network Smoothness for Algorithmic Neural Reasoning
TL;DR: In this article , the authors investigate the underlying hypothesis in the most simple conceivable scenario -the addition of real numbers and find that two layer neural networks fail to learn the structure of this task and growing the network's width leads to a complex division of input space.
References
More filters
Proceedings Article
Adam: A Method for Stochastic Optimization
Diederik P. Kingma,Jimmy Ba +1 more
TL;DR: This work introduces Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments, and provides a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework.
Book
Bayesian learning for neural networks
TL;DR: Bayesian Learning for Neural Networks shows that Bayesian methods allow complex neural network models to be used without fear of the "overfitting" that can occur with traditional neural network learning methods.
Journal ArticleDOI
A Unifying View of Sparse Approximate Gaussian Process Regression
TL;DR: A new unifying view, including all existing proper probabilistic sparse approximations for Gaussian process regression, relies on expressing the effective prior which the methods are using, and highlights the relationship between existing methods.
Journal Article
In Defense of One-Vs-All Classification
Ryan Rifkin,Aldebaro Klautau +1 more
TL;DR: It is argued that a simple "one-vs-all" scheme is as accurate as any other approach, assuming that the underlying binary classifiers are well-tuned regularized classifiers such as support vector machines.
Proceedings Article
Gaussian processes for Big data
TL;DR: In this article, the authors introduce stochastic variational inference for Gaussian process models, which enables the application of Gaussian Process (GP) models to data sets containing millions of data points.