Piecewise linear activations substantially shape the loss surfaces of neural networks.
Fengxiang HeBohan WangDacheng TaoPublished in: ICLR (2020)
Keyphrases
- piecewise linear
- neural network
- principal curves
- smooth curves
- dynamic programming
- chaotic map
- pattern recognition
- finite sets
- surface shape
- shape model
- shape analysis
- free form surfaces
- regression algorithm
- genetic algorithm
- surface registration
- shape space
- hyperplane
- shape descriptors
- surface orientation
- shape representation
- shape matching
- range data
- high dimensional
- surface model
- medial axis
- surface reconstruction
- image processing