Stabilizing gradients for deep neural networks via efficient SVD parameterization

Mathematics

• Recurrent Neural Networks 100%
• Singular value decomposition 85%
• Parameterization 78%
• Neural Networks 73%
• Gradient 59%
• Transition Matrix 43%
• Training 41%
• Singular Values 40%
• Multilayer Perceptron 28%
• Reflector 26%
• Numerical Linear Algebra 26%
• Parameterise 24%
• Range of data 23%
• Banks 23%
• Orthogonal matrix 23%
• Demonstrate 23%
• Expressive Power 22%
• Generalization 19%
• Experimental Results 19%
• Parametrization 19%
• Spectrality 14%
• Converge 12%
• Framework 11%

Engineering & Materials Science

• Parameterization 82%
• Recurrent neural networks 80%
• Singular value decomposition 80%
• Deep neural networks 75%
• Linear algebra 24%
• Multilayer neural networks 20%