On the Computational Power of Online Gradient Descent

  • 2018-07-03 16:56:14
  • Vaggos Chatziafratis, Tim Roughgarden, Joshua R. Wang
  • 22

Abstract

We prove that the evolution of weight vectors in online gradient descent canencode arbitrary polynomial-space computations, even in the special case ofsoft-margin support vector machines. Our results imply that, under weakcomplexity-theoretic assumptions, it is impossible to reason efficiently aboutthe fine-grained behavior of online gradient descent.

 

Introduction (beta)

None

 

Conclusion (beta)

None