Deep learning generalizes because the parameter-function map is biased towards simple functions

Mathematics

• Simple Function 100%
• Deep Neural Network 100%
• Deep Learning Method 100%
• Regularization 33%
• Error Bound 16%
• Parameter Space 16%
• Gaussian Process 16%
• Marginal Likelihood 16%
• Algorithmic Information Theory 16%
• Bayes Approach 16%
• Function Space 16%
• World Problems 16%
• Training Set 16%
• Output Function 16%
• Boolean Function 16%
• Connected Network 16%
• Likelihood 16%
• Intrinsic Property 16%