@inproceedings{verwimp-etal-2018-state,
title = "State Gradients for {RNN} Memory Analysis",
author = "Verwimp, Lyan and
Van hamme, Hugo and
Renkens, Vincent and
Wambacq, Patrick",
editor = "Linzen, Tal and
Chrupa{\l}a, Grzegorz and
Alishahi, Afra",
booktitle = "Proceedings of the 2018 {EMNLP} Workshop {B}lackbox{NLP}: Analyzing and Interpreting Neural Networks for {NLP}",
month = nov,
year = "2018",
address = "Brussels, Belgium",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/W18-5443",
doi = "10.18653/v1/W18-5443",
pages = "344--346",
abstract = "We present a framework for analyzing what the state in RNNs remembers from its input embeddings. We compute the gradients of the states with respect to the input embeddings and decompose the gradient matrix with Singular Value Decomposition to analyze which directions in the embedding space are best transferred to the hidden state space, characterized by the largest singular values. We apply our approach to LSTM language models and investigate to what extent and for how long certain classes of words are remembered on average for a certain corpus. Additionally, the extent to which a specific property or relationship is remembered by the RNN can be tracked by comparing a vector characterizing that property with the direction(s) in embedding space that are best preserved in hidden state space.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="verwimp-etal-2018-state">
<titleInfo>
<title>State Gradients for RNN Memory Analysis</title>
</titleInfo>
<name type="personal">
<namePart type="given">Lyan</namePart>
<namePart type="family">Verwimp</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hugo</namePart>
<namePart type="family">Van hamme</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vincent</namePart>
<namePart type="family">Renkens</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Patrick</namePart>
<namePart type="family">Wambacq</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2018-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tal</namePart>
<namePart type="family">Linzen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Grzegorz</namePart>
<namePart type="family">Chrupała</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Afra</namePart>
<namePart type="family">Alishahi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Brussels, Belgium</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>We present a framework for analyzing what the state in RNNs remembers from its input embeddings. We compute the gradients of the states with respect to the input embeddings and decompose the gradient matrix with Singular Value Decomposition to analyze which directions in the embedding space are best transferred to the hidden state space, characterized by the largest singular values. We apply our approach to LSTM language models and investigate to what extent and for how long certain classes of words are remembered on average for a certain corpus. Additionally, the extent to which a specific property or relationship is remembered by the RNN can be tracked by comparing a vector characterizing that property with the direction(s) in embedding space that are best preserved in hidden state space.</abstract>
<identifier type="citekey">verwimp-etal-2018-state</identifier>
<identifier type="doi">10.18653/v1/W18-5443</identifier>
<location>
<url>https://aclanthology.org/W18-5443</url>
</location>
<part>
<date>2018-11</date>
<extent unit="page">
<start>344</start>
<end>346</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T State Gradients for RNN Memory Analysis
%A Verwimp, Lyan
%A Van hamme, Hugo
%A Renkens, Vincent
%A Wambacq, Patrick
%Y Linzen, Tal
%Y Chrupała, Grzegorz
%Y Alishahi, Afra
%S Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP
%D 2018
%8 November
%I Association for Computational Linguistics
%C Brussels, Belgium
%F verwimp-etal-2018-state
%X We present a framework for analyzing what the state in RNNs remembers from its input embeddings. We compute the gradients of the states with respect to the input embeddings and decompose the gradient matrix with Singular Value Decomposition to analyze which directions in the embedding space are best transferred to the hidden state space, characterized by the largest singular values. We apply our approach to LSTM language models and investigate to what extent and for how long certain classes of words are remembered on average for a certain corpus. Additionally, the extent to which a specific property or relationship is remembered by the RNN can be tracked by comparing a vector characterizing that property with the direction(s) in embedding space that are best preserved in hidden state space.
%R 10.18653/v1/W18-5443
%U https://aclanthology.org/W18-5443
%U https://doi.org/10.18653/v1/W18-5443
%P 344-346
Markdown (Informal)
[State Gradients for RNN Memory Analysis](https://aclanthology.org/W18-5443) (Verwimp et al., EMNLP 2018)
ACL
- Lyan Verwimp, Hugo Van hamme, Vincent Renkens, and Patrick Wambacq. 2018. State Gradients for RNN Memory Analysis. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 344–346, Brussels, Belgium. Association for Computational Linguistics.