Deep Neural Networks Learn Non-Smooth Functions Effectively

We theoretically discuss why deep neural networks (DNNs) performs better than other models in some cases by investigating statistical properties of DNNs for non-smooth functions. While DNNs have empirically shown higher performance than other standard methods, understanding its mechanism is still a challenging problem. From an aspect of the statistical theory, it is known many standard methods att

[elu_18]

深層学習を統計理論で解析した論文がAISTATSに採択されました。データが生成されてる状況について、深層学習が理論上の最適性能を達成するが、他手法（カーネル法など）が最適性を持たない状況を特定することで、現状