@inproceedings{bengio:2015:nips,
  author = {S. Bengio and O. Vinyals and N. Jaitly and N. Shazeer},
  title = {Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks},
  booktitle = {Advances In Neural Information Processing Systems, {NIPS}},
  web = {http://arxiv.org/abs/1506.03099},
  year = 2015,
  url = {publications/ps/bengio_2015_nips.ps.gz},
  pdf = {publications/pdf/bengio_2015_nips.pdf},
  djvu = {publications/djvu/bengio_2015_nips.djvu},
  original = {2015/sequence_sampling/v3},
  abstract = {Recurrent Neural Networks can be trained to produce sequences of tokens given some input, as exemplified by recent results in machine translation and image captioning. The current approach to training them consists of maximizing the likelihood of each token in the sequence given the current (recurrent) state and the previous token. At inference, the unknown previous token is then replaced by a token generated by the model itself. This discrepancy between training and inference can yield errors that can accumulate quickly along the generated sequence.  We propose a curriculum learning strategy to gently change the training process from a fully guided scheme using the true previous token, towards a less guided scheme which mostly uses the generated token instead.  Experiments on several sequence prediction tasks show that this approach yields significant improvements. Moreover, it was used successfully in our winning entry to the MSCOCO image captioning challenge, 2015.},
  categorie = {C},
}

{"_id":"GxCZRFckaYD3Hk7ae","bibbaseid":"bengio-vinyals-jaitly-shazeer-scheduledsamplingforsequencepredictionwithrecurrentneuralnetworks-2015","downloads":0,"creationDate":"2017-05-10T16:03:41.197Z","title":"Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks","author_short":["Bengio, S.","Vinyals, O.","Jaitly, N.","Shazeer, N."],"year":2015,"bibtype":"inproceedings","biburl":"http://bengio.abracadoudou.com/samy.bib","bibdata":{"bibtype":"inproceedings","type":"inproceedings","author":[{"firstnames":["S."],"propositions":[],"lastnames":["Bengio"],"suffixes":[]},{"firstnames":["O."],"propositions":[],"lastnames":["Vinyals"],"suffixes":[]},{"firstnames":["N."],"propositions":[],"lastnames":["Jaitly"],"suffixes":[]},{"firstnames":["N."],"propositions":[],"lastnames":["Shazeer"],"suffixes":[]}],"title":"Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks","booktitle":"Advances In Neural Information Processing Systems, NIPS","web":"http://arxiv.org/abs/1506.03099","year":"2015","url":"publications/ps/bengio_2015_nips.ps.gz","pdf":"publications/pdf/bengio_2015_nips.pdf","djvu":"publications/djvu/bengio_2015_nips.djvu","original":"2015/sequence_sampling/v3","abstract":"Recurrent Neural Networks can be trained to produce sequences of tokens given some input, as exemplified by recent results in machine translation and image captioning. The current approach to training them consists of maximizing the likelihood of each token in the sequence given the current (recurrent) state and the previous token. At inference, the unknown previous token is then replaced by a token generated by the model itself. This discrepancy between training and inference can yield errors that can accumulate quickly along the generated sequence. We propose a curriculum learning strategy to gently change the training process from a fully guided scheme using the true previous token, towards a less guided scheme which mostly uses the generated token instead. Experiments on several sequence prediction tasks show that this approach yields significant improvements. Moreover, it was used successfully in our winning entry to the MSCOCO image captioning challenge, 2015.","categorie":"C","bibtex":"@inproceedings{bengio:2015:nips,\n author = {S. Bengio and O. Vinyals and N. Jaitly and N. Shazeer},\n title = {Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks},\n booktitle = {Advances In Neural Information Processing Systems, {NIPS}},\n web = {http://arxiv.org/abs/1506.03099},\n year = 2015,\n url = {publications/ps/bengio_2015_nips.ps.gz},\n pdf = {publications/pdf/bengio_2015_nips.pdf},\n djvu = {publications/djvu/bengio_2015_nips.djvu},\n original = {2015/sequence_sampling/v3},\n abstract = {Recurrent Neural Networks can be trained to produce sequences of tokens given some input, as exemplified by recent results in machine translation and image captioning. The current approach to training them consists of maximizing the likelihood of each token in the sequence given the current (recurrent) state and the previous token. At inference, the unknown previous token is then replaced by a token generated by the model itself. This discrepancy between training and inference can yield errors that can accumulate quickly along the generated sequence. We propose a curriculum learning strategy to gently change the training process from a fully guided scheme using the true previous token, towards a less guided scheme which mostly uses the generated token instead. Experiments on several sequence prediction tasks show that this approach yields significant improvements. Moreover, it was used successfully in our winning entry to the MSCOCO image captioning challenge, 2015.},\n categorie = {C},\n}\n\n","author_short":["Bengio, S.","Vinyals, O.","Jaitly, N.","Shazeer, N."],"key":"bengio:2015:nips","id":"bengio:2015:nips","bibbaseid":"bengio-vinyals-jaitly-shazeer-scheduledsamplingforsequencepredictionwithrecurrentneuralnetworks-2015","role":"author","urls":{"Paper":"http://bengio.abracadoudou.com/publications/ps/bengio_2015_nips.ps.gz"},"downloads":0},"search_terms":["scheduled","sampling","sequence","prediction","recurrent","neural","networks","bengio","vinyals","jaitly","shazeer"],"keywords":[],"authorIDs":[],"dataSources":["9NCW2CDr4M3s5DvNX"]}