Optimal Completion Distillation for Sequence Learning
We present Optimal Completion Distillation (OCD), a training procedure for optimizing sequence to sequence models based on edit distance. OCD is efficient, has no hyper-parameters of its own, and does not require pretraining or joint optimization with conditional log-likelihood. Given a partial sequence generated by the model, we first identify the set of optimal suffixes that minimize the total edit distance, using an efficient dynamic programming algorithm. Then, for each position of the generated sequence, we use a target distribution that puts equal probability on the first token of all the optimal suffixes. OCD achieves the state-of-the-art performance on end-to-end speech recognition, on both Wall Street Journal and Librispeech datasets, achieving WER and WER respectively.
Recent advances in natural language processing and speech recognition hinge on the development of expressive neural network architectures for sequence to sequence (seq2seq) learning (Sutskever et al., 2014; Bahdanau et al., 2015). Such encoder-decoder architectures are adopted in both machine translation (Bahdanau et al., 2015; Wu et al., 2016; Hassan et al., 2018) and speech recognition systems (Chan et al., 2016; Bahdanau et al., 2016a; Chiu et al., 2017) to achieve impressive performance beyond traditional multi-stage pipelines (Koehn et al., 2007; Povey et al., 2011). Improving the building blocks of seq2seq models can fundamentally advance machine translation and speech recognition, and positively impact other domains such as image captioning (Xu et al., 2015), parsing (Vinyals et al., 2015), summarization (Rush et al., 2015), and program synthesis (Zhong et al., 2017).
To improve the key components of seq2seq models, one can either design better architectures, or develop better learning algorithms. Recent architectures using convolution (Gehring et al., 2017) and self attention (Vaswani et al., 2017) have proved to be useful, especially to facilitate efficient training. On the other hand, despite many attempts to mitigate the limitations of Maximum Likelihood Estimation (MLE) (Ranzato et al., 2016; Wiseman and Rush, 2016; Norouzi et al., 2016; Bahdanau et al., 2017; Leblond et al., 2018), MLE is still considered the dominant approach for training seq2seq models. Current alternative approaches require pre-training or joint optimization with conditional log-likelihood. They are difficult to implement and require careful tuning of new hyper-parameters (e.g. mixing ratios). In addition, alternative approaches typically do not offer a substantial performance improvement over a well tuned MLE baseline, especially when label smoothing (Pereyra et al., 2017; Edunov et al., 2018) or scheduled sampling (Bengio et al., 2015) is used.
In this paper, we borrow ideas from search-based structured prediction (Daumé et al., 2009; Ross et al., 2011)
and policy distillation (Rusu et al., 2016) and develop an efficient algorithm for optimizing seq2seq models
based on edit distance
We always train on prefixes generated by sampling from the model that is being optimized.
For each generated prefix, we identify all optimal suffixes that result in a minimum total edit distance v.s. the ground truth target using an efficient dynamic programming algorithm.
We teach the model to optimally extend each generated prefix by maximizing the average log probability of the first tokens of the optimal suffixes identified in step 2.
The proposed OCD algorithm is efficient, straightforward to implement, and has no tunable hyper-parameters of its own. Our key contributions include:
We propose OCD, a stand-alone algorithm for optimizing seq2seq models based on edit distance. OCD is scalable to real-world datasets with long sequences and large vocabularies, and consistently outperforms Maximum Likelihood Estimation (MLE) by a large margin.
Given a target sequence of length and a generated sequence of length , we present an algorithm that identifies all of the optimal extensions for each prefix of the generated sequence.
We demonstrate the effectiveness of OCD on end-to-end speech recognition using attention-based seq2seq models. On the Wall Street Journal dataset, OCD achieves a Character Error Rate (CER) of and a Word Error Rate (WER) of without language model rescoring, outperforming all prior work (Table 4). On Librispeech, OCD achieves state-of-the-art WER of on “test-clean” and on “test-other” sets (Table 5).
2 Background: Sequence Learning with MLE
Given a dataset of input output pairs , we are interested in learning a mapping from an input to a target output sequence . Let denote the set of all sequences of tokens from a finite vocabulary with variable but finite lengths. Often learning a mapping is formulated as optimizing the parameters of a conditional distribution . Then, the final sequence prediction under the probabilistic model is performed by exact or approximate inference (e.g. via beam search) as:
Similar to the use of log loss for supervised classification, the standard approach to optimize the parameters of the conditional probabilistic model entails maximizing a conditional log-likelihood objective:
This approach to learning the parameters is called Maximum Likelihood Estimation (MLE).
Sutskever et al. (2014) propose the use of recurrent neural networks (RNNs) for autoregressive seq2seq modeling to tractably optimize . An autoregressive model estimates the conditional probability of the target sequence given the source one token at a time, often from left-to-right. A special end-of-sequence token is appended at the end of all of target sequences to handle variable length. The conditional probability of given is decomposed via the chain rule as,
where denotes a prefix of the sequence . To estimate the probability of a token given a prefix and an input , denoted , different architectures have been proposed. Some papers (e.g. Britz et al. (2017)) have investigated the use of LSTM (Hochreiter and Schmidhuber, 1997) and GRU (Cho et al., 2014) cells, while others proposed new architectural components such as soft attention (Bahdanau et al., 2015), convolution (Gehring et al., 2017), and self-attention (Vaswani et al., 2017). This paper focuses on the objective function used for optimizing autoregressive seq2seq models and presents a technique applicable to any autoregressive architecture.
2.1 Limitations of MLE for Autoregressive Models
In order to maximize the conditional log-likelihood (2) of an autoregressive seq2seq model (3), one provides the model with a prefix of tokens from the ground truth target sequence, denoted , and maximizes the log-probability of as the next token. This resembles a teacher walking a student through a sequence of perfect decisions, where the student learns as a passive observer. However, during inference one uses beam search (1), wherein the student needs to generate each token by conditioning on its own previous outputs, i.e. instead of . This creates a discrepancy between training and test known as exposure bias (Ranzato et al., 2016). Appendix B expands this further.
Concretely, we highlight two limitations with the use of MLE for autoregressive seq2seq modeling:
There is a mismatch between the prefixes seen by the model during training and inference. When the distribution of is different from the distribution of , then the student will find themselves in a novel situation that they have not been trained for. This can result in poor generalization, especially when the training set is small or the model size is large.
There is a mismatch between the training loss and the task evaluation metric. During training, one optimizes the log-probability of the ground truth output sequence, which is often different from the task evaluation metric (e.g. edit distance for speech recognition).
There has been a recent surge of interest in understanding and mitigating the limitations of MLE for autoregressive seq2seq modeling. In Section 4 we discuss prior work in detail after presenting our approach below.
3 Optimal Completion Distillation
To alleviate the mismatch between inference and training, we never train on ground truth target sequences. Instead, we always train on sequences generated by sampling from the current model that is being optimized. Let denote a sequence generated by the current model, and denote the ground truth target. Applying MLE to autoregressive models casts the problem of sequence learning as optimizing a mapping from ground truth prefixes to correct next tokens. By contrast, the key question that arises when training on model samples is the choice of targets for learning a similar mapping from generated prefixes to next tokens. Instead of using a set of pre-specified targets, OCD solves a prefix-specific problem to find optimal extensions that lead to the best completions according to the task evaluation metric. Then, OCD encourages the model to extend each prefix with the set of optimal choices for the next token.
Our notion of optimal completion depends on the task evaluation metric denoted , which measures the similarity between two complete sequences, e.g. the ground truth target v.s. a generated sequence. Edit distance is a common task metric. Our goal in sequence learning is to train a model, which achieves high scores of . Drawing connection with the goal of reinforcement learning (Sutton and Barto, 1998), let us recall the notion of optimal Q-values. Optimal Q-values for a state-action pair , denoted , represent the maximum future reward that an agent can accumulate after taking an action at a state by following with optimal subsequent actions. Similarly, we define Q-values for a prefix and the extending token , as the maximum score attainable by concatenating with an optimal suffix to create a full sequence . Formally,
Then, the optimal extension for a prefix can be defined as tokens that attain the maximal Q-values, i.e. . This formulation allows for a prefix to be sampled on-policy from the model , or drawn off-policy in any way. Table 1 includes an example ground truth target from the Wall Street Journal dataset and the corresponding generated sample from a model. We illustrate that for some prefixes there exist more than a single optimal extension leading to the same edit distance.
Given -values for our prefix-token pairs, we use an exponential transform followed by normalization to convert Q-values to a soft optimal distribution over the next token:
where is a temperature parameter. Note that resembles the label smoothing parameter helpful within MLE. In our experiments, we used the limit of resulting in hard targets and no hyper-parameter tuning.
In order to learn the optimal distribution over sequences, the OCD objective minimizes a KL divergence between and the model’s distribution, , over the extensions at every time step ,
We minimize this loss for all input-output pairs and their corresponding model samples. To compute for every , we still need to calculate the Q-values. For the important class of sequence learning problems where edit distance is the evaluation metric, we develop a dynamic programming algorithm to calculate Q-values exactly and efficiently for all prefixes of a sequence , discussed below.
|Target sequence||a s _ h e _ t a l k s _ h i s _ w i f e|
|Generated sequence||a s _ e e _ t a l k s _ w h o s e _ w i f e|
|Optimal extensions for||a s _ h e _ t a l k s _ h i i s _ _ w i f e|
|edit distance (OCD||h h i w|
3.1 Optimal Q-values for Edit Distance
We propose a dynamic programming algorithm to calculate optimal Q-values exactly and efficiently for the reward metric of negative edit distance, i.e. . Given two sequences and , we compute the Q-values for every prefix and any extending token with an asymptotic complexity of . Assuming that , our algorithm does not increase the time complexity over MLE, since computing the cross-entropy losses in MLE also requires a complexity of .
|Edit Distance Table||OCD Targets|
|T||3||2||2||2||3||4||4||U, N, D||2|
|R||4||3||3||3||3||4||5||U, N, D, A||3|
Recall the Levenshtein algorithm (Levenshtein, 1966) for calculating the minimum number of edits (insertion, deletion and substitution) required to convert sequences and to each other based on,
Table 2 shows an example edit distance table for sequences “Satrapy” and “Sunday”. Our goal is to identify the set of all optimal suffixes that result in a full sequences with a minimum edit distance v.s. .
The edit distance resulting from any potential suffix is lower bounded by ,
Let’s consider the path that traces back to connecting each cell to an adjacent parent cell, which provides the minimum value among the three options in (7). Such a path for tracing edit distance between “Satrapy” and “Sunday” is shown in Table 2. Suppose the path crosses row at a cell . Since the operations in (7) are non-decreasing, the edit distance along the path cannot decrease, so . ∎
Then, consider any such that . Let denote a suffix of . We conclude that , because on the one hand there is a particular edit path that results in edits, and on the other hand is a lower bound according to Lemma 1. Hence any such is an optimal suffix for . Further, it is straightforward to prove by contradiction that the set of optimal suffixes is limited to suffixes corresponding to .
Since the set of optimal completions for is limited to , the only extensions that can lead to maximum reward are the starting token of such suffixes (). Since as well, we can identify the optimal extensions by calculating the edit distances between all prefixes of and all prefixes of which can be efficiently calculated by dynamic programming in . For a prefix after we calculate the minimum edit distance among all prefixes of , we set the for all where has edit distance equal to . We set the for any other token to . We provide the details of our modified Levenshtein algorithm to efficiently compute the for all and in Appendix A.
4 Related Work
Our work is inspired by Imitation Learning techniques (Ross et al., 2011; Ross and Bagnell, 2014; Sun et al., 2018), where a student policy is optimized to imitate an expert teacher. DAgger (Ross et al., 2011) in particular is closely related, where a dataset of trajectories are collected from past student models, and a cost-sensitive classification problem is optimized. By contrast, in OCD, the exploration policy is drawn directly from an online student and a KL loss is optimized. AggreVaTeD (Sun et al., 2017) assumes access to an unbiased estimate of Q-values and relies on techniques such as conjugate gradient and variance reduction regularization. OCD gains access to exact Q-values and uses regular SGD for optimization. Importantly, our roll-in prefixes are drawn only from the model, and we do not require mixing in ground truth (a.k.a. expert, teacher) samples. Cheng and Boots (2018) showed that mixing in ground truth samples is an essential regularizer for value aggregation convergence in imitation learning.
More closely related to our work is Policy Distillation (Rusu et al., 2016), where a deep Q-Network (DQN) agent (Mnih et al., 2015) is optimized as the teacher first. Then, action sequences are sampled from the teacher and the learned Q-value estimates are distilled (Hinton et al., 2014) into a smaller student network using a KL loss. OCD adopts a similar loss function, but rather than estimating Q-values using bootstrapping, we estimate exact -values using dynamic programming. Moreover, we draw samples from the student rather than the teacher.
The learning to search (L2S) literature (Daumé III and Marcu, 2005; Chang et al., 2015; Goodman et al., 2016; Leblond et al., 2018) also attempts at estimating the cost of each action by examining multiple roll-outs of a generated prefix. SeaRNN (Leblond et al., 2018) approximates the cost of each token by computing the task loss for vocabulary size roll-outs at each time step. This makes it difficult to scale to real world datasets, where either the sequences are long or the vocabulary is large. We exploit the special structure in edit distance and find exact Q-values efficiently. Unlike L2S and SeaRNN, which require ground truth prefixes to stabilize training, we solely train on model samples.
Scheduled Sampling (SS) (Bengio et al., 2015) and Data as Demonstrator (DaD) (Venkatraman et al., 2015) are similar techniques that aim to alleviate the mismatch between training and inference prefix distributions. SS generates prefixes by substituting some of the ground truth tokens with samples from the model via a tunable mixing schedule. SS uses the ground truth sequence as the training targets uninformed of the synthetic prefixes. By contrast, OCD draws tokens via sampling from the model, without any schedule. While SS can only handle substitution errors, OCD considers insertions and deletions as well by solving an alignment problem to find optimal targets.
Approaches based on reinforcement-Learning (RL) have also been applied to sequence prediction problems, including REINFORCE (Ranzato et al., 2016), Actor-Critic (Bahdanau et al., 2017) and Self-critical Sequence Training (Rennie et al., 2017). These methods sample sequences from the model’s distribution and backpropagate a sequence-level task objective (e.g. edit distance). Beam Search Optimization (Wiseman and Rush, 2016) and Edit-based Minimum Bayes Risk (EMBR) (Prabhavalkar et al., 2018) is similar, except the sampling procedure is replaced with beam search. These training methods suffer from high variances and credit assignment problems. By contrast, OCD takes advantage of the decomposition of the sequence-level objective into token level optimal completion targets. This reduces the variance of the gradient and stabilizes the model. Crucially, unlike most RL-based approaches, we neither need MLE pretraining or joint optimization with log-likelihood. Bahdanau et al. (2016b) also notice some of the nice structure of edit distance, but they optimize the model by regressing its outputs to edit distance values leading to suboptimal performance. Rather, we first construct the optimal policy and then use knowledge distillation for training.
Reward Augmented Maximum Likelihood (RAML) (Norouzi et al., 2016) and its variants (Ma et al., 2017; Elbayad et al., 2018; Wang et al., 2018) are also similiar to RL-based approaches. Instead of sampling from the model’s distribution, RAML samples sequences from the true exponentiated reward distribution. However, sampling from the true distribution is often difficult and intractable. RAML suffers from the same problems as RL-based methods in credit assignment.
Generally, OCD excels at training from scratch, which makes it an ideal substitution for MLE. Hence, OCD is orthogonal to methods which require MLE pretraining or joint optimization.
We conduct our experiments on speech recogntion on the Wall Street Journal (WSJ) (Paul and Baker, 1992) and Librispeech (Panayotov et al., 2015) benchmarks. We only compare end-to-end speech recognition approaches that do not incorporate language model rescoring. On both WSJ and Librispeech, our proposed OCD (Optimal Completion Distillation) algorithm significantly outperforms our own strong baselines including MLE (Maximum Likelihood Estimation with label smoothing) and SS (scheduled sampling with a well-tuned schedule). Moreover, OCD significantly outperforms all prior work, achieving a new state-of-the-art on two competitive benchmarks.
5.1 Wall Street Journal
The WSJ dataset is readings of three separate years of the Wall Street Journal. We use the standard configuration of si for training, dev for validation and report both test Character Error Rate (CER) and Word Error Rate (WER) on eval. We tokenize the dataset to English characters and punctuation. Our model is an attention-based seq2seq network with a deep convolutional frontend as used in Zhang et al. (2017). During inference, we use beam search with a beam size of for all of our models. We describe the architecture and hyperparameter details in Appendix C. We first analyze some key characteristics of the OCD model separately, and then compare our results with other baselines and state-of-the-art methods.
Training prefixes and generalization. We emphasize that during training, the generated prefixes sampled from the model do not match the ground truth sequence, even at the end of training. We define OCD prefix mismatch as the fraction of OCD training tokens that do not match corresponding ground truth training tokens at each position. Assuming that the generated prefix sequence is perfectly matched with the ground truth sequence, then the OCD targets would simply be the following tokens of the ground truth sequence. Hence, OCD becomes equivalent to MLE. Figure 1 shows that OCD prefixes mismatch is more than for the most of the training. This suggests that OCD and MLE are training on very different input prefix trajectories. Further, Figure 2 depicts validation CER as a function of training CER for different model checkpoints during training, where we use beam search on both training and validation sets to obtain CER values. Even at the same training CER, we observe better validation error for OCD, which suggests that OCD improves generalization of MLE, possibly because OCD alleviates the mismatch between training and inference.
Impact of edit distance. We further investigate the role of the optimizer by experimenting with different losses. Table 3 compares the test CER and WER of the schedule sampling with a fixed probability schedule of and OCD model. Both of the models are trained only on sampled trajectories. The main difference is their optimizers, where the SS model is optimizing the log likelihood of ground truth (a.k.a. Hamming distance). The significant drop in CER of SS emphasizes the necessity of pretraining or joint training with MLE for models such as SS. OCD is trained from random initialization and does not require MLE pretraining, nor does it require joint optimization with MLE. We also emphasize that unlike SS, we do not need to tune an exploration schedule, OCD prefixes are simply always sampled from the model from the start of training. We note that even fine tuning a pre-trained SS model which achieves CER with sampling increases the CER to %. This emphasizes the importance of making the loss a function of the model input prefixes, as opposed to the ground truth prefixes. Appendix D covers another aspect of optimizing Edit distance rather than Hamming distance.
Target distribution. Another baseline which is closer to MLE framework is selecting only one correct target. Table 3 compares OCD with several Optimal Completion Target (OCT) models. In OCT, we optimize the log-likelihood of one target, which at each step we pick dynamically based on the minimum edit distance completion similar to OCD. We experiment with several different strategies when there is more than one character that can lead to minimum CER. In the OCT (Shortest), we select the token that would minimize the CER and the final length of the sequence. In the OCT (Same #Words), we select the token that in addition to minimum CER, would lead to the closest number of words to the target sequence. We show that OCD achieves significantly better CER and WER over the other optimization strategies compared in Table 3. This highlights the importance of optimizing for the entire set of optimal completion targets, as opposed to a single target.
|Optimal Completion Target (Shortest)||3.8||12.7|
|Optimal Completion Target (Same #Words)||3.3||10.2|
|Optimal Completion Distillation||3.1||9.3|
|CTC (Graves and Jaitly; 2014)||9.2||30.1|
|CTC + REINFORCE (Graves and Jaitly; 2014)||8.4||27.3|
|Gram-CTC (Liu et al.; 2017)||-||16.7|
|seq2seq (Bahdanau et al.; 2016a)||6.4||18.6|
|seq2seq + TLE (Bahdanau et al.; 2016b)||5.9||18.0|
|seq2seq + LS (Chorowski and Jaitly; 2017)||-||10.6|
|seq2seq + CNN (Zhang et al.; 2017)||-||10.5|
|seq2seq + LSD (Chan et al.; 2017)||-||9.6|
|seq2seq + CTC (Kim et al.; 2017)||7.4||-|
|seq2seq + TwinNet (Serdyuk et al.; 2018)||6.2||-|
|seq2seq + MLE + REINFORCE (Tjandra et al.; 2018)||6.1||-|
|seq2seq + MLE||3.6||10.6|
|seq2seq + SS||3.6||10.2|
|seq2seq + OCD||3.1||9.3|
State-of-the-art. Our model trained with OCD optimizes for CER; we achieve 3.1% CER and 9.3% WER, substantially outperforming our baseline by 14% relatively on CER and 12% relatively on WER. In terms of CER, our work substantially outperforms prior work as compared in Table 4, with the closest being Tjandra et al. (2018) trained with policy gradients on CER. In terms of WER, our work is also outperforming Chan et al. (2017), which uses subword units while our model emits characters.
For the Librispeech dataset, we train on the full training set (960h audio data) and validate our results on the dev-other set. We report the results both on the “clean” and “other” test set. We use Byte Pair Encoding (BPE) (Sennrich et al., 2016) for the output token segmentation. BPE token set is an open vocabulary set since it includes the characters as well as common words and n-grams. We use 10k BPE tokens and report both CER and WER as the evaluation metric. We describe the architecture and hyperparameter details in Appendix C.
Our MLE baseline achieves WER, while OCD achieves WER on test-clean ( improvement) and improves the state-of-the-art results over Zeyer et al. (2018). test-other is the more challenging test split ranked by the WER of a model trained on WSJ (Panayotov et al., 2015) mainly because readers accents deviate more from US-English accents. On test-other our MLE baseline achieves 15.4%, while our OCD model achieves 13.3% WER, outperforming the 15.4% WER of Zeyer et al. (2018). Table 5 compares our results with other recent works and the MLE baseline on Librispeech.
|Wav2letter (Collobert et al., 2016)||6.9||7.2||-||-|
|Gated ConvNet (Liptchinsky et al., 2017)||-||6.7||-||20.8|
|Cold Fusion (Sriram et al., 2018)||3.9||7.5||9.3||17.0|
|Invariant Representation Learning (Liang et al., 2018)||3.3||-||11.0||-|
|Pretraining+seq2seq+CTC (Zeyer et al., 2018)||-||4.9||-||15.4|
|seq2seq + MLE||2.9||5.7||8.4||15.4|
|seq2seq + OCD||1.7||4.5||6.4||13.3|
This paper presents Optimal Completion Distillation (OCD), a training procedure for optimizing autoregressive sequence models base on edit distance. OCD is applicable to on-policy or off-policy trajectories, and in this paper, we demonstrate its effectiveness on samples drawn from the model in an online fashion. Given any prefix, OCD creates an optimal extension policy by computing the exact optimal Q-values via dynamic programming. The optimal extension policy is distilled by minimizing a KL divergence between the optimal policy and the model. OCD does not require MLE initialization or joint optimization with conditional log-likelihood. OCD achieves % CER and % WER on the competitive WSJ speech recognition task, and WER on Librispeech without any language model. OCD outperforms all published work on end-to-end speech recognition, including our own well-tuned MLE and scheduled sampling baselines without introducing new hyper-parameters.
Appendix A OCD Algorithm
Complexity. The total time complexity for calculating the sequence loss using OCD is where is the vocabulary size and is the sequence length. MLE loss has a time complexity of for calculating the softmax loss at each step. Therefore, assuming that OCD does not change the time complexity compared to the baseline seq2seq+MLE. The memory cost of the OCD algorithm is , for the dynamic programming in line 4 - line 13 of Proc. 1 and for storing the stepwise values. MLE also stores the one-hot encoding of targets at each step with a cost of . Therefore, the memory complexity does not change compared to the MLE baseline either.
Although the loss calculation has the same complexity as MLE, online sampling from the model to generate the input of next RNN cell (as in OCD and SS) is generally slower than reading the ground truth (as in MLE). Therefore, overall a naive implementation of OCD is slower than our baseline MLE in terms of number of step time. However, since OCD is stand alone and can be trained off-policy, we can also train on stale samples and untie the input generation worker from the training workers. In this case it is as fast as the MLE baseline.
Run through. As an example of how this algorithm works, consider the sequence “SUNDAY” as reference and “SATURDAY” as hypothesis. Table A.1 first shows how to extract optimal targets and their respective -values from the table of edit distances between all prefixes of reference and all prefixes of hypothesis. At each row highlighted cells indicate the prefixes which has minimum edit distance in the row. The next character at these indices are the Optimal targets for that row. At each step the -value for the optimal targets is negative of the minimum edit distance and for the non-optimal characters it is one smaller.
Table A.1 also illustrates how appending the optimal completions for the prefix “SA” of the hypothesis can lead to the minimum total edit distance. Concatenating with both reference suffixes, “UNDAY” and “NDAY” will result in an edit distance of . Therefore, predicting “U” or “N” at step 2 can lead to the maximum attainable reward of .
|Edit Distance||OCD Targets||Q-values|
|T||3||2||2||2||3||4||4||U, N, D||-2|
Appendix B Exposure bias
A key limitation of teacher forcing for sequence learning stems from the discrepancy between the training and test objectives. One trains the model using conditional log-likelihood , but evaluates the quality of the model using empirical reward .
|(a) Teacher Forcing (MLE)||(b) Scheduled Sampling|
|(c) Policy Gradient||(d) Optimal Completion Distillation|
Unlike teacher forcing and Scheduled Sampling (SS), policy gradient approaches (e.g. Ranzato et al. , Bahdanau et al. ) and OCD aim to optimize the empirical reward objective (4) on the training set. We illustrate four different training strategies of MLE, SS, Policy Gradient and OCD in Figure B.1. The drawback of policy gradient techniques is twofold: 1) they cannot easily incorporate ground truth sequence information except through the reward function, and 2) they have difficulty reducing the variance of the gradients to perform proper credit assignment. Accordingly, most policy gradient approaches Ranzato et al. , Bahdanau et al. , Wu et al.  pre-train the model using teacher forcing. By contrast, the OCD method proposed in this paper defines an optimal completion policy for any off-policy prefix by incorporating the ground truth information. Then, OCD optimizes a token level log-loss and alleviates the credit assignment problem. Finally, training is much more stable, and we do not require initialization nor joint optimization with MLE.
There is an intuitive notion of exposure bias Ranzato et al.  discussed in the literature as a limitation of teacher forcing. We formalize this notion as follows. One can think of the optimization of the log loss (2) in an autoregressive models as a classification problem, where the input to the classifier is a tuple and the correct output is , where . Then the training dataset comprises different examples and different prefixes of the ground truth sequence. The key challenge is that once the model is trained, one should not expect the model to generalize to a new prefix that does not come from the training distribution of . This problem can become severe as becomes more dissimilar to correct prefixes. During inference, when one conducts beam search with a large beam size then one is more likely to discover wrong generalization of , because the sequence is optimized globally. A natural strategy to remedy this issue is to train on arbitrary prefixes. Unlike the aforementioned techniques OCD can train on any prefix given its off-policy nature.
Figure B.2 illustrates how increasing the beam size for MLE and SS during inference decreases their performance on WSJ datasets to above WER. OCD suffers a degradation in the performance too but it never gets above WER.
Appendix C Architecture
WSJ. The input audio signal is converted into -dimensional filterbank features computed every ms with delta and delta-delta acceleration, normalized with per-speaker mean and variance generated by Kaldi [Povey et al., 2011]. Our encoder uses 2-layers of convolutions with filters, stride and channels, followed by a convolutional LSTM with 1D-convolution of filter width , followed by 3 LSTM layers with cell size. We also apply batch-normalization between each layer in the encoder. The attention-based decoder is a 1-layer LSTM with cell size with content-based attention. We train our models for 300 epochs of batch size 8 with 8 async workers. We separately tune the learning rate for our baseline and OCD model, for OCD vs for baseline. We apply a single drop of learning rate when validation CER plateaus, the same as for our baseline. We implemented our experiments
Librispeech. Since the dataset is larger than WSJ, we use a larger batch size of 16, smaller learning rate of for baseline and for OCD. We remove the convolutional LSTM layers of the encoder, increase the number of LSTM layers in the encoder to 6, and increase the LSTM cell size to 384. All other configs are the same as the WSJ setup.
Appendix D Hamming distance VS Edit Distance during training
Figure D.3 plots the edit distance on training data of OCD and MLE for fixed hamming distances during training. The plot shows that for a fixed Hamming distance (which is the metric that MLE correlates with more), OCD achieves a lower edit distance compared to MLE. This gives evidence that OCD is indeed optimizing for edit distance as intended.
- Edit distance between two sequences and is the minimum number of insertion, deletion, and substitution edits required to convert to and vice versa.
- We are in the process of releasing the code for OCD.
- Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: a system for large-scale machine learning. In OSDI, volume 16, pages 265–283, 2016.
- Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. ICLR, 2015.
- Dzmitry Bahdanau, Jan Chorowski, Dmitriy Serdyuk, Philemon Brakel, and Yoshua Bengio. End-to-End Attention-based Large Vocabulary Speech Recognition. ICASSP, 2016a.
- Dzmitry Bahdanau, Dmitriy Serdyuk, Philemon Brakel, Nan Rosemary Ke, Jan Chorowski, Aaron Courville, and Yoshua Bengio. Task Loss Estimation for Sequence Prediction. ICLR Workshop, 2016b.
- Dzmitry Bahdanau, Philemon Brakel, Kelvin Xu, Anirudh Goyal, Ryan Lowe, Joelle Pineau, Aaron Courville, and Yoshua Bengio. An Actor-Critic Algorithm for Sequence Prediction. ICLR, 2017.
- Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam M. Shazeer. Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks. NIPS, 2015.
- Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc Le. Massive exploration of neural machine translation architectures. arXiv:1703.03906, 2017.
- William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. Listen, Attend and Spell: A Neural Network for Large Vocabulary Conversational Speech Recognition. ICASSP, 2016.
- William Chan, Yu Zhang, and Navdeep Jaitly. Latent Sequence Decompositions. ICLR, 2017.
- Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agarwal, Hal Daumé III, and John Langford. Learning to Search Better Than Your Teacher. In ICML, 2015.
- Ching-An Cheng and Byron Boots. Convergence of value aggregation for imitation learning. arXiv preprint arXiv:1801.07292, 2018.
- Chung-Cheng Chiu, Tara N Sainath, Yonghui Wu, Rohit Prabhavalkar, Patrick Nguyen, Zhifeng Chen, Anjuli Kannan, Ron J Weiss, Kanishka Rao, Katya Gonina, et al. State-of-the-art speech recognition with sequence-to-sequence models. arXiv:1712.01769, 2017.
- Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. EMNLP, 2014.
- Jan Chorowski and Navdeep Jaitly. Towards better decoding and language model integration in sequence to sequence models. INTERSPEECH, 2017.
- Ronan Collobert, Christian Puhrsch, and Gabriel Synnaeve. Wav2letter: an end-to-end convnet-based speech recognition system. arXiv preprint arXiv:1609.03193, 2016.
- Hal Daumé, III, John Langford, and Daniel Marcu. Search-based structured prediction. Mach. Learn. J., 2009.
- Hal Daumé III and Daniel Marcu. Learning as search optimization: Approximate large margin methods for structured prediction. ICML, 2005.
- Sergey Edunov, Myle Ott, Michael Auli, David Grangier, and Marc’Aurelio Ranzato. Classical structured prediction losses for sequence to sequence learning. NAACL, 2018.
- Maha Elbayad, Laurent Besacier, and Jakob Verbeek. Token-level and sequence-level loss smoothing for rnn language models. ACL, 2018.
- Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional Sequence to Sequence Learning. In arXiv:1705.03122, 2017.
- James Goodman, Andreas Vlachos, and Jason Naradowsky. Noise reduction and targeted exploration in imitation learning for abstract meaning representation parsing. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 1–11, 2016.
- Alex Graves and Navdeep Jaitly. Towards End-to-End Speech Recognition with Recurrent Neural Networks. ICML, 2014.
- Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan Clark, Christian Federmann, Xuedong Huang, Marcin Junczys-Dowmunt, William Lewis, Mu Li, et al. Achieving human parity on automatic chinese to english news translation. arXiv:1803.05567, 2018.
- Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In Neural Information Processing Systems: Workshop Deep Learning and Representation Learning Workshop, 2014.
- Sepp Hochreiter and Jurgen Schmidhuber. Long Short-Term Memory. Neural Comput., 9, 1997.
- Suyoun Kim, Takaaki Hori, and Shinji Watanabe. Joint CTC-Attention based End-to-End Speech Recognition using Multi-task Learning. ICASSP, 2017.
- Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. Moses: Open source toolkit for statistical machine translation. Proceedings of the 45th annual meeting of the ACL on interactive poster and demonstration sessions, 2007.
- Rémi Leblond, Jean-Baptiste Alayrac, Anton Osokin, and Simon Lacoste-Julien. SEARNN: Training RNNs with global-local losses. ICLR, 2018.
- Vladimir I Levenshtein. Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10(8):707–710, 1966.
- Davis Liang, Zhiheng Huang, and Zachary C Lipton. Learning noise-invariant representations for robust speech recognition. arXiv preprint arXiv:1807.06610, 2018.
- Vitaliy Liptchinsky, Gabriel Synnaeve, and Ronan Collobert. Letter-Based Speech Recognition with Gated ConvNets. In arXiv:1712.09444, 2017.
- Hairong Liu, Zhenyao Zhu, Xiangang Li, and Sanjeev Satheesh. Gram-CTC: Automatic Unit Selection and Target Decomposition for Sequence Labelling. ICML, 2017.
- Xuezhe Ma, Pengcheng Yin, Jingzhou Liu, Graham Neubig, and Eduard Hovy. Softmax q-distribution estimation for structured prediction: A theoretical interpretation for raml. arXiv:1705.07136, 2017.
- Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 2015.
- Mohammad Norouzi, Samy Bengio, Zhifeng Chen, Navdeep Jaitly, Mike Schuster, Yonghui Wu, and Dale Schuurmans. Reward Augmented Maximum Likelihood for Neural Structured Prediction. NIPS, 2016.
- Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on, pages 5206–5210. IEEE, 2015.
- Douglas B Paul and Janet M Baker. The design for the wall street journal-based csr corpus. In Proceedings of the workshop on Speech and Natural Language, pages 357–362. Association for Computational Linguistics, 1992.
- Gabriel Pereyra, George Tucker, Jan Chorowski, Łukasz Kaiser, and Geoffrey Hinton. Regularizing Neural Networks by Penalizing Confident Output Distributions. ICLR Workshop, 2017.
- Daniel Povey, Arnab Ghoshal, Gilles Boulianne, et al. The Kaldi Speech Recognition Toolkit. ASRU, 2011.
- Rohit Prabhavalkar, Tara N. Sainath, Yonghui Wu, Patrick Nguyen, Zhifeng Chen, Chung-Cheng Chiu, and Anjuli Kannan. Minimum Word Error Rate Training for Attention-based Sequence-to-Sequence Models. In ICASSP, 2018.
- Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence Level Training with Recurrent Neural Networks. ICLR, 2016.
- Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jarret Ross, and Vaibhava Goel. Self-critical Sequence Training for Image Captioning. In CVPR, 2017.
- Stephane Ross and J Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. arXiv preprint arXiv:1406.5979, 2014.
- Stephane Ross, Geoffrey J Gordon, and J Andrew Bagnell. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. AISTATS, 2011.
- Alexander M. Rush, Sumit Chopra, and Jason Weston. A Neural Attention Model for Abstractive Sentence Summarization. In EMNLP, 2015.
- Andrei A. Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy Distillation. ICLR, 2016.
- Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural Machine Translation of Rare Words with Subword Units. ACL, 2016.
- Dmitriy Serdyuk, Nan Rosemary Ke, Alessandro Sordoni, Adam Trischler, Chris Pal, and Yoshua Bengio. Twin Networks: Matching the Future for Sequence Generation. ICLR, 2018.
- Anuroop Sriram, Heewoo Jun, Sanjeev Satheesh, and Adam Coates. Cold fusion: Training seq2seq models together with language models. In INTERSPEECH, 2018.
- Wen Sun, Arun Venkatraman, Geoffrey J Gordon, Byron Boots, and J Andrew Bagnell. Deeply aggrevated: Differentiable imitation learning for sequential prediction. arXiv preprint arXiv:1703.01030, 2017.
- Wen Sun, J Andrew Bagnell, and Byron Boots. Truncated horizon policy search: Combining reinforcement learning & imitation learning. arXiv preprint arXiv:1805.11240, 2018.
- Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to Sequence Learning with Neural Networks. NIPS, 2014.
- Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction. MIT Press, 1998.
- Andros Tjandra, Sakriani Sakti, and Satoshi Nakamura. Sequence-to-Sequence ASR Optimization via Reinforcement Learning. ICASSP, 2018.
- Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017.
- Arun Venkatraman, Martial Hebert, and J Andrew Bagnell. Improving multi-step prediction of learned time series models. In AAAI, 2015.
- Oriol Vinyals, Lukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey Hinton. Grammar as a Foreign Language. NIPS, 2015.
- Xinyi Wang, Hieu Pham, Zihang Dai, and Graham Neubig. Switchout: an efficient data augmentation algorithm for neural machine translation. EMNLP, 2018.
- Sam Wiseman and Alexander M. Rush. Sequence-to-Sequence Learning as Beam-Search Optimization. EMNLP, 2016.
- Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv:1609.08144, 2016.
- Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation with Visual Attention. ICML, 2015.
- Albert Zeyer, Kazuki Irie, Ralf Schlüter, and Hermann Ney. Improved training of end-to-end attention models for speech recognition. arXiv preprint arXiv:1805.03294, 2018.
- Yu Zhang, William Chan, and Navdeep Jaitly. Very Deep Convolutional Networks for End-to-End Speech Recognition. ICASSP, 2017.
- Victor Zhong, Caiming Xiong, and Richard Socher. Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning. arXiv:1709.00103, 2017.