Implicitly Defined Neural Networks for Sequence Labeling
Abstract
In this paper, we propose an implicit neural network architecture and show that it can be computed in a reasonably efficient manner. Our architecture relaxes the causality assumption in formulating recurrent neural networks, so that the hidden states of the network are coupled together, in order to improve performance on complex, long-range dependencies in either direction of a sequence. We contrast our architecture with a bidirectional RNN, and show that our proposed architecture the bidirectional network matches its performance on one task, while providing an ensembling benefit greater than ensembling multiple bidirectional networks.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 31, 2017
- Accession Number
- AD1032195
Entities
People
- Brian J. Thompson
- Michaeel M. Kazi
Organizations
- Massachusetts Institute of Technology