Deep Perm-Set Net: Learn to Predict Sets with Unknown Permutation and Cardinality Using Deep Neural Networks
We present a novel approach for learning to predict sets with unknown permutation and cardinality using deep neural networks. Even though the output of many real-world problems, e.g. object detection, are naturally expressed as sets of entities, existing deep learning architectures hinder a trivial extension to deal with this unstructured output. Even deep architectures that handle sequential data, such as recurrent neural networks, can only output an ordered set and may not guarantee a valid solution, i.e. a set with unique elements. In this paper, we derive a mathematical formulation for set prediction using feed-forward neural networks, where the output has unknown and unfixed cardinality and permutation. Specifically, in our formulation we incorporate the permutation as unobservable variable and estimate its distribution during the learning process using alternating optimization. We demonstrate the validity of this formulation on two relevant problems including object detection and a complex CAPTCHA test.
READ FULL TEXT