WHAT HAPPENS AFTER SGD REACHES ZERO LOSS? -A MATHEMATICAL FRAMEWORK

Keyphrases

• Mathematical Framework 100%
• Stochastic Gradient Descent 100%
• Zero Loss 100%
• Loss Function 33%
• Gradient Descent 33%
• Label Noise 33%
• Noise Covariance 33%
• Implicit Bias 33%
• Learning Rate 16%
• Global Analysis 16%
• Stochastic Differential Equations 16%
• Deep Learning 16%
• Regularizer 16%
• 2-step 16%
• Local Analysis 16%
• Full Characterization 16%
• Local Minimizer 16%
• Gradient Noise 16%
• Track Gradient 16%
• Minimax Optimal 16%
• Regularization Effect 16%

Mathematics

• Stochastics 100%
• Manifold 50%
• Loss Function 33%
• Upper Bound 33%
• Covariance 33%
• Linear Models 16%
• Regularization 16%
• Minimax 16%
• Stochastic Differential Equation 16%
• Like Term 16%
• Local Minimizer 16%
• Deep Learning Method 16%

Computer Science

• Stochastic Gradient Descent 100%
• Gradient Descent 33%
• Implicit Bias 33%
• Regularization 16%
• Stochastic Differential 16%
• Learning Rate 16%
• Local Analysis 16%
• Global Analysis 16%
• Deep Learning Method 16%