Architecture for Predicting Sets

Having an effective model for predicting sets would be useful in tasks such as object detection, image tagging, forming a team of employees in an office and so on, where the output we are interested in, is a set (order invariant). Trying to predict sets using multi-class classification type of techniques face the issue of unknown cardinality, along with its difficulty in capturing dependencies between different elements of the set during prediction. Another approach is to use models that predict sequences to predict sets. In this case we have to come up with a model that can find the right sequence out of the many possible sequences that can constitute a set, as some orders are easy to model than the others. We propose a recurrent neural networks based architecture for predicting sets, which uses an order invariant loss function to learn the sequence of prediction automatically.

Authors

• Janarthanan Rajendran
• Satinder Singh

Keywords

No keywords are indexed for this paper.