Information Geometry of Orthogonal Initializations and Training

Abstract

Recently mean field theory has been successfully used to analyze properties of wide, random neural networks. It gave rise to a prescriptive theory for initializing feed-forward neural networks with orthogonal weights, which ensures that both the forward propagated activations and the backpropagated gradients are near ${\ell }_2$ isometries and as a consequence training is orders of magnitude faster. Despite strong empirical performance, the mechanisms by which critical initializations confer an advantage in the optimization of deep neural networks are poorly understood. Here we show a novel connection between the maximum curvature of the optimization landscape (gradient smoothness) as measured by the Fisher information matrix (FIM) and the spectral radius of the input-output Jacobian, which partially explains why more isometric networks can train much faster. Furthermore, given that orthogonal weights are necessary to ensure that gradient norms are approximately preserved at initialization, we experimentally investigate the benefits of maintaining orthogonality throughout training, from which we conclude that manifold optimization of weights performs well regardless of the smoothness of the gradients. Moreover, motivated by experimental results we show that a low condition number of the FIM is not predictive of faster learning.

arXiv comment10 pages and 5 figures; 5 page appendix

References in corpus (20)

• Adam: A Method for Stochastic Optimization
• Neural Tangent Kernel: Convergence and Generalization in Neural Networks
• Exact solutions to the nonlinear dynamics of learning in deep linear neural networks
• Deep Neural Networks as Gaussian Processes
• Optimizing Neural Networks with Kronecker-factored Approximate Curvature
• How Does Batch Normalization Help Optimization?
• Unitary Evolution Recurrent Neural Networks
• Parseval Networks: Improving Robustness to Adversarial Examples
• On orthogonality and learning recurrent networks with long term dependencies
• The Emergence of Spectral Universality in Deep Networks
• Dynamical Isometry and a Mean Field Theory of RNNs: Gating Enables Signal Propagation in Recurrent Neural Networks
• Recurrent Orthogonal Networks and Long-Memory Tasks
• A Mean Field Theory of Batch Normalization
• Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach
• Generalized BackPropagation, Étude De Cas: Orthogonality
• Optimization on Submanifolds of Convolution Kernels in CNNs
• Practical Gauss-Newton Optimisation for Deep Learning
• All You Need is Beyond a Good Init: Exploring Better Solution for Training Extremely Deep Convolutional Neural Networks with Orthonormality and Modulation
• Exponential convergence rates for Batch Normalization: The power of length-direction decoupling in non-convex optimization
• Deep Information Propagation

Cited by in corpus (4)

• Sorting out Lipschitz function approximation
• A Comprehensive and Modularized Statistical Framework for Gradient Norm Equality in Deep Neural Networks
• Wide-minima Density Hypothesis and the Explore-Exploit Learning Rate Schedule
• On the Neural Tangent Kernel of Deep Networks with Orthogonal Initialization