On the Interpretation of Recurrent Neural Networks as Finite State Machines. Oliva, C. & Lago-Fernández, L. F. In Tetko, I. V., Kůrková, V., Karpov, P., & Theis, F., editors, Artificial Neural Networks and Machine Learning – ICANN 2019: Theoretical Neural Computation, of Lecture Notes in Computer Science, pages 312–323, Cham, 2019. Springer International Publishing. doi abstract bibtex The behavior of simple recurrent neural networks trained on regular languages is analyzed in terms of accuracy and interpretability. We use controlled amounts of noise and L1 regularization to obtain stable and accurate responses that are at the same time highly interpretable, and introduce a shocking mechanism that reactivates silent neurons when learning stops due to an excessive regularization. Proper parameter tuning allows the networks to develop a strong generalization capacity, and at the same time provides solutions that may be interpreted as finite automata. Experiments carried out with different regular languages show that, in all cases, the trained networks display activation patterns that automatically cluster into a set of discrete states without any need to explicitly perform quantization. Analysis of the transitions between states in response to the input symbols reveals that the networks are in fact implementing a finite state machine that in all cases matches the regular expressions used to generate the training data.
@inproceedings{oliva_interpretation_2019,
address = {Cham},
series = {Lecture {Notes} in {Computer} {Science}},
title = {On the {Interpretation} of {Recurrent} {Neural} {Networks} as {Finite} {State} {Machines}},
isbn = {978-3-030-30487-4},
doi = {10.1007/978-3-030-30487-4_25},
abstract = {The behavior of simple recurrent neural networks trained on regular languages is analyzed in terms of accuracy and interpretability. We use controlled amounts of noise and L1 regularization to obtain stable and accurate responses that are at the same time highly interpretable, and introduce a shocking mechanism that reactivates silent neurons when learning stops due to an excessive regularization. Proper parameter tuning allows the networks to develop a strong generalization capacity, and at the same time provides solutions that may be interpreted as finite automata. Experiments carried out with different regular languages show that, in all cases, the trained networks display activation patterns that automatically cluster into a set of discrete states without any need to explicitly perform quantization. Analysis of the transitions between states in response to the input symbols reveals that the networks are in fact implementing a finite state machine that in all cases matches the regular expressions used to generate the training data.},
language = {en},
booktitle = {Artificial {Neural} {Networks} and {Machine} {Learning} – {ICANN} 2019: {Theoretical} {Neural} {Computation}},
publisher = {Springer International Publishing},
author = {Oliva, Christian and Lago-Fernández, Luis F.},
editor = {Tetko, Igor V. and Kůrková, Věra and Karpov, Pavel and Theis, Fabian},
year = {2019},
pages = {312--323},
}
Downloads: 0
{"_id":"ERhfjGnLEnYW6o6NS","bibbaseid":"oliva-lagofernndez-ontheinterpretationofrecurrentneuralnetworksasfinitestatemachines-2019","author_short":["Oliva, C.","Lago-Fernández, L. F."],"bibdata":{"bibtype":"inproceedings","type":"inproceedings","address":"Cham","series":"Lecture Notes in Computer Science","title":"On the Interpretation of Recurrent Neural Networks as Finite State Machines","isbn":"978-3-030-30487-4","doi":"10.1007/978-3-030-30487-4_25","abstract":"The behavior of simple recurrent neural networks trained on regular languages is analyzed in terms of accuracy and interpretability. We use controlled amounts of noise and L1 regularization to obtain stable and accurate responses that are at the same time highly interpretable, and introduce a shocking mechanism that reactivates silent neurons when learning stops due to an excessive regularization. Proper parameter tuning allows the networks to develop a strong generalization capacity, and at the same time provides solutions that may be interpreted as finite automata. Experiments carried out with different regular languages show that, in all cases, the trained networks display activation patterns that automatically cluster into a set of discrete states without any need to explicitly perform quantization. Analysis of the transitions between states in response to the input symbols reveals that the networks are in fact implementing a finite state machine that in all cases matches the regular expressions used to generate the training data.","language":"en","booktitle":"Artificial Neural Networks and Machine Learning – ICANN 2019: Theoretical Neural Computation","publisher":"Springer International Publishing","author":[{"propositions":[],"lastnames":["Oliva"],"firstnames":["Christian"],"suffixes":[]},{"propositions":[],"lastnames":["Lago-Fernández"],"firstnames":["Luis","F."],"suffixes":[]}],"editor":[{"propositions":[],"lastnames":["Tetko"],"firstnames":["Igor","V."],"suffixes":[]},{"propositions":[],"lastnames":["Kůrková"],"firstnames":["Věra"],"suffixes":[]},{"propositions":[],"lastnames":["Karpov"],"firstnames":["Pavel"],"suffixes":[]},{"propositions":[],"lastnames":["Theis"],"firstnames":["Fabian"],"suffixes":[]}],"year":"2019","pages":"312–323","bibtex":"@inproceedings{oliva_interpretation_2019,\n\taddress = {Cham},\n\tseries = {Lecture {Notes} in {Computer} {Science}},\n\ttitle = {On the {Interpretation} of {Recurrent} {Neural} {Networks} as {Finite} {State} {Machines}},\n\tisbn = {978-3-030-30487-4},\n\tdoi = {10.1007/978-3-030-30487-4_25},\n\tabstract = {The behavior of simple recurrent neural networks trained on regular languages is analyzed in terms of accuracy and interpretability. We use controlled amounts of noise and L1 regularization to obtain stable and accurate responses that are at the same time highly interpretable, and introduce a shocking mechanism that reactivates silent neurons when learning stops due to an excessive regularization. Proper parameter tuning allows the networks to develop a strong generalization capacity, and at the same time provides solutions that may be interpreted as finite automata. Experiments carried out with different regular languages show that, in all cases, the trained networks display activation patterns that automatically cluster into a set of discrete states without any need to explicitly perform quantization. Analysis of the transitions between states in response to the input symbols reveals that the networks are in fact implementing a finite state machine that in all cases matches the regular expressions used to generate the training data.},\n\tlanguage = {en},\n\tbooktitle = {Artificial {Neural} {Networks} and {Machine} {Learning} – {ICANN} 2019: {Theoretical} {Neural} {Computation}},\n\tpublisher = {Springer International Publishing},\n\tauthor = {Oliva, Christian and Lago-Fernández, Luis F.},\n\teditor = {Tetko, Igor V. and Kůrková, Věra and Karpov, Pavel and Theis, Fabian},\n\tyear = {2019},\n\tpages = {312--323},\n}\n\n\n\n\n\n\n\n\n\n\n\n","author_short":["Oliva, C.","Lago-Fernández, L. F."],"editor_short":["Tetko, I. V.","Kůrková, V.","Karpov, P.","Theis, F."],"key":"oliva_interpretation_2019","id":"oliva_interpretation_2019","bibbaseid":"oliva-lagofernndez-ontheinterpretationofrecurrentneuralnetworksasfinitestatemachines-2019","role":"author","urls":{},"metadata":{"authorlinks":{}},"downloads":0,"html":""},"bibtype":"inproceedings","biburl":"https://bibbase.org/zotero/warren.pettine","dataSources":["zYjFbyFmZWKpRCD4j"],"keywords":[],"search_terms":["interpretation","recurrent","neural","networks","finite","state","machines","oliva","lago-fernández"],"title":"On the Interpretation of Recurrent Neural Networks as Finite State Machines","year":2019}