|タイトル||Tropical geometry of Deep Neural Networks|
2018年11月13日 10:45-11:45 (+ 30min)
|場所||慶應義塾大学理工学部 14棟 7階733 (Building 14, Room 733)|
|内容||We give a review on the paper “Tropical Geometry of Deep Neural Networks”, which is published in ICML 2018.
They establish a connection between feedforward neural networks with ReLU activation function and Tropical geometry, which is a field in Mathematics to treat piecewise linear functions. Precisely speaking, they show that the family of such neural networks is equivalent to the family of tropical rational maps.
We start with a brief introduction to Tropical geometry and see their result.