Training Input-Output Recurrent Neural Networks through Spectral Methods
- Creators
- Sedghi, Hanie
- Anandkumar, Anima
Abstract
We consider the problem of training input-output recurrent neural networks (RNN) for sequence labeling tasks. We propose a novel spectral approach for learning the network parameters. It is based on decomposition of the cross-moment tensor between the output and a non-linear transformation of the input, based on score functions. We guarantee consistent learning with polynomial sample and computational complexity under transparent conditions such as non-degeneracy of model parameters, polynomial activations for the neurons, and a Markovian evolution of the input sequence. We also extend our results to Bidirectional RNN which uses both previous and future information to output the label at each time point, and is employed in many NLP tasks such as POS tagging.
Additional Information
The authors thank Majid Janzamin for discussions on sample complexity and constructive comments on the draft. We thank Ashish Sabharwal for editorial comments on the draft. This work was done during the time H. Sedghi was a visiting researcher at University of California, Irvine and was supported by NSF Career award FG15890. A. Anandkumar is supported in part by Microsoft Faculty Fellowship, NSF Career award CCF-1254106, ONR award N00014-14-1-0665, ARO YIP award W911NF-13-1-0084, and AFOSR YIP award FA9550-15-1-0221.Attached Files
Submitted - 1603.00954.pdf
Files
Name | Size | Download all |
---|---|---|
md5:e0328244a100c6a76697f7715478a213
|
371.3 kB | Preview Download |
Additional details
- Eprint ID
- 94325
- Resolver ID
- CaltechAUTHORS:20190401-123315920
- FG15890
- NSF
- Microsoft Faculty Fellowship
- CCF-1254106
- NSF
- N00014-14-1-0665
- Office of Naval Research (ONR)
- W911NF-13-1-0084
- Army Research Office (ARO)
- FA9550-15-1-0221
- Air Force Office of Scientific Research (AFOSR)
- Created
-
2019-04-01Created from EPrint's datestamp field
- Updated
-
2023-06-02Created from EPrint's last_modified field