Abstract

A novel Neural Network architecture is proposed using the mathematically andphysically rich idea of vector fields as hidden layers to perform nonlineartransformations in the data. The data points are interpreted as particlesmoving along a flow defined by the vector field which intuitively representsthe desired movement to enable classification. The architecture moves the datapoints from their original configuration to anew one following the streamlinesof the vector field with the objective of achieving a final configuration whereclasses are separable. An optimization problem is solved through gradientdescent to learn this vector field.

 

Quick Read (beta)