Input Convex Neural Networks

• DOI: 10.48550/arxiv.1609.07152
• Publication Date: 2016-01-01
• Authors: Brandon Amos; Lei Xu; J. Zico Kolter
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Optimization and Control (math.OC); 0202 electrical engineering, electronic engineering, information engineering; FOS: Mathematics; 02 engineering and technology; Mathematics - Optimization and Control; Machine Learning (cs.LG)

This paper presents the input convex neural network architecture. These are scalar-valued (potentially deep) networks with constraints on parameters such that output of is a function (some of) inputs. The allow for efficient inference via optimization over some inputs to given others, and can be applied settings including structured prediction, data imputation, reinforcement learning, others. In this we lay basic groundwork these models, proposing methods inference, analyze their representational power. We show many existing architectures made input-convex minor modification, develop specialized algorithms tailored setting. Finally, highlight performance multi-label image completion, learning problems, where improvement state art in cases.