Input-Convex Deep Networks

Brandon Amos, J. Zico Kolter

Abstract:

This paper introduces a new class of neural networks that we
refer to as input-convex neural networks, networks that are convex in
their inputs (as opposed to their parameters). We discuss the nature and
representational power of these networks, illustrate how the prediction
(inference) problem can be solved via convex optimization, and discuss their
application to structured prediction problems. We highlight a few simple
examples of these networks applied to classification tasks, where we illustrate
that the networks perform substantially better than any other approximator we
are aware of that is convex in its inputs.

Conflicts: cmu.edu