[MXNET-1438] New Loss Function for Neural Information Retrieval

• Type: New Feature
• Status: To Do
• Priority: Major
• Resolution: Unresolved
• Component/s: Gluon
• Assignee: Unassigned
• Reporter: Anjishnu
• Created: 19/Nov/19 23:50
• Updated: 26/Jul/20 05:24

Description

Currently MxNet implements only a few loss functions useful for neural information retrieval - in particular the Triplet Loss. In our recent research at Amazon we showed that this loss is not ideal for neural information retrieval settings and devise a new loss function which significantly outperformed Triplet Loss while improving training time by a factor of 10.

Relevant Paper: https://arxiv.org/abs/1905.12786 (Published at EMNLP 2019 - WNUT)

We would like to implement this as a loss function in Gluon to help any other people working on neural information retrieval systems using MxNet.

Issue Links

• links to: GitHub Pull Request #17298