Exact solutions to the nonlinear dynamics of learning in deep linear neural networks
By Andrew M. Saxe, James L. McClelland, Surya Ganguli. 2013
2,139 citations recorded by the source corpus.
Despite the widespread practical success of deep learning methods, our theoretical understanding of the dynamics of learning in deep neural networks remains quite sparse. We attempt to bridge the gap between the theory and practice of deep learning by systematically analyzing learning dynamics for the restricted case of deep linear neural networks. Despite the linearity of their input-output map, such networks have nonlinear gradient descent dynamics on weights that change with the addition of each new hidden layer. We show that deep linear networks exhibit nonlinear learning phenomena similar to those seen in simulations of nonlinear networks, including long plateaus followed by rapid transitions to lower error solutions, and faster convergence from greedy unsupervised pretraining initial conditions than from random initial conditions. We provide an analytical description of these phenomena by finding new exact solutions to the nonlinear dynamics of deep learning. Our theoretical analysis also reveals the surprising finding that as the depth of a network approaches infinity, learning speed can nevertheless remain finite: for a special class of initial conditions on the weights, very deep networks incur only a finite, depth independent, delay in learning speed relative to shallow networks. We show that, under certain conditions on the training data, unsupervised pretraining can find this special class of initial conditions, while scaled random Gaussian initializations cannot. We further exhibit a new class of random orthogonal initial conditions on weights that, like unsupervised pre-training, enjoys depth independent learning times. We further show that these initial conditions also lead to faithful propagation of gradients even in deep nonlinear networks, as long as they operate in a special regime known as the edge of chaos.
Subfields: Artificial Intelligence, Statistical and Nonlinear Physics, Optimization & Training, Scaling & Learning Theory
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Yoshua Bengio
- Mohit Bansal
- Sergey Levine
- Jianfeng Gao
- Quoc V. Le
- Percy Liang
- Ruslan Salakhutdinov
- Kyunghyun Cho
In Statistical and Nonlinear Physics
- Greg Yang
- Yoshua Bengio
- Sergey Levine
- Sam Schoenholz
- Roger Baker Grosse
- Yuanzhi Li
- Jeffrey Pennington
- Pieter Abbeel
In Optimization & Training
- Yoshua Bengio
- Quoc V. Le
- Soham De
- Roger Baker Grosse
- Razvan Pascanu
- Simon Lacoste-Julien
- Stefano Soatto
- Yann Dauphin
In Scaling & Learning Theory
- Greg Yang
- Yuanzhi Li
- Tengyu Ma
- Simon Shaolei Du
- Jason D. Lee
- Sam Schoenholz
- Yoshua Bengio
- Matthieu Wyart
Citations in the corpus
- Deep Residual Learning for Image Recognition (2015, background): “This problem, however, has been largely addressed by normalized initialization [23, 9, 37, 13] and intermediate normalization layers [16], which enable networks with tens of layers to start converging for stochastic gradient descent (SGD) with backpropagation [22].”
- Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation (2014, methodology): “For the recurrent weight matrices, we first sampled from a white Gaussian distribution and used its left singular vectors matrix, following (Saxe et al., 2014).”
- Highway Networks (2015, background): “Optimization of deep networks has proven to be considerably more difficult, leading to research on initialization schemes (Glorot & Ben-gio, 2010; Saxe et al., 2013; He et al., 2015), techniques of training networks in multiple stages (Simonyan & Zis-serman, 2014; Romero et al., 2014) or with…”
- Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification (2015, background): “Despite the prevalence of rectifier networks, recent improvements of models [37, 28, 12, 29, 33] and theoretical guidelines for training them [8, 27] have rarely focused on the properties of the rectifiers.”
- Very Deep Convolutional Networks for Large-Scale Image Recognition (2014, neutral): “by adjusting the random initialisation procedure [15].”
- Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (2017, background): “A number of prior works have explored the topic of sensitivity in deep networks, often in the context of studying initialization techniques for weight matrices (Saxe et al., 2014; Kirkpatrick et al., 2016).”
- Gaussian Error Linear Units (GELUs) (2016, background): “Weights are initialized with unit norm rows, as this has positive impact on each nonlinearity’s performance (Hendrycks & Gimpel, 2016; Mishkin & Matas, 2016; Saxe et al., 2014).”
- Wide neural networks of any depth evolve as linear models under gradient descent (2019, background): “[18] study the training dynamics of deep linear networks, in which the nonlinearities are treated as identity functions.”
- Scaling description of generalization with number of parameters in deep learning (2019, methodology): “The weights of the network are initialized according to the random orthogonal scheme [45] and all biases are initialized to zero.”
- Gradient Descent Happens in a Tiny Subspace (2018, background): “5Other works where the dynamics of gradient descent were analyzed directly include Fukumizu; Saxe et al. (2013); Arora et al. (2018).”
- Large Scale GAN Training for High Fidelity Natural Image Synthesis (2018, methodology): “Both G and D networks are initialized with Orthogonal Initialization (Saxe et al., 2014).”
- WaveGrad: Estimating Gradients for Waveform Generation (2020, methodology): “Orthogonal initialization (Saxe et al., 2014) is used for all UBlocks and DBlocks.”
- Learning Layout and Style Reconfigurable GANs for Controllable Image Synthesis (2020, methodology): “Parameters of the Generator and the Discriminator are initialized using the Orthogonal Initialization method [55].”
- Implicit Regularization of Discrete Gradient Dynamics in Deep Linear Neural Networks (2019, neutral): “Prior work [Saxe et al., 2013, 2014, Advani and Saxe, 2017, Saxe et al., 2018, Lampinen and Ganguli, 2019] studied the gradient dynamics of two-layer linear networks and proved a result similar to our Thm. 2.”