Learning Longer Memory in
Recurrent Neural Networks
Recurrent neural network is a powerful model that learns temporal patterns in sequential data. For a long time, it was believed that recurrent networks are difficult to train using simple optimizers, such as stochastic gradient descent, due to the so-called vanishing gradient problem. In this paper, we show that learning longer term patterns in real data, such as in natural language, is perfectly possible using gradient descent. This is achieved by using a slight structural modification of the simple recurrent neural network architecture. We encourage some of the hidden units to change their state slowly by making part of the recurrent weight matrix close to identity, thus forming a kind of longer term memory. We evaluate our model on language modeling tasks on benchmark datasets, where we obtain similar performance to the much more complex Long Short Term Memory (LSTM) networks (Hochreiter & Schmidhuber, 1997).
|Tomas Mikolov, Armand Joulin, Sumit Chopra, Michael Mathieu & Marc’Aurelio Ranzato|
|Facebook Artificial Intelligence Research|
|New York City, NY, 10003, USA|
Models of sequential data, such as natural language, speech and video, are the core of many machine learning applications. This has been widely studied in the past with approaches taking their roots in a variety of fields (Goodman, 2001b; Young et al., 1997; Koehn et al., 2007). In particular, models based on neural networks have been very successful recently, obtaining state-of-the-art performances in automatic speech recognition (Dahl et al., 2012), language modeling (Mikolov, 2012) and video classification (Simonyan & Zisserman, 2014). These models are mostly based on two families of neural networks: feedforward neural networks and recurrent neural networks.
Feedforward architectures such as time-delayed neural networks usually represent time explicitly with a fixed-length window of the recent history (Rumelhart et al., 1985). While this type of models work well in practice, fixing the window size makes long-term dependency harder to learn and can only be done at the cost of a linear increase of the number of parameters.
The recurrent architectures, on the other hand, represent time recursively. For example, in the simple recurrent network (SRN) (Elman, 1990), the state of the hidden layer at a given time is conditioned on its previous state. This recursion allows the model to store complex signals for arbitrarily long time periods, as the state of the hidden layer can be seen as the memory of the model. In theory, this architecture could even encode a “perfect” memory by simply copying the state of the hidden layer over time.
While theoretically powerful, these recurrent models were widely considered to be hard to train due to the so-called vanishing and exploding gradient problems (Hochreiter, 1998; Bengio et al., 1994). Mikolov (2012) showed how to avoid the exploding gradient problem by using simple, yet efficient strategy of gradient clipping. This allowed to efficiently train these models on large datasets by using only simple tools such as stochastic gradient descent and back-propagation through time (Williams & Zipser, 1995; Werbos, 1988).
Nevertheless, simple recurrent networks still suffer from the vanishing gradient problem: as gradients are propagated back through time, their magnitude will almost always exponentially shrink. This makes memory of the SRNs focused only on short term patterns, practically ignoring longer term dependencies. There are two reasons why this happens. First, standard nonlinearities such as the sigmoid function have a gradient which is close to zero almost everywhere. This problem has been partially solved in deep networks by using the rectified linear units (ReLu) (Nair & Hinton, 2010). Second, as the gradient is backpropagated through time, its magnitude is multiplied over and over by the recurrent matrix. If the eigenvalues of this matrix are small (i.e., less than one), the gradient will converge to zero rapidly. Empirically, gradients are usually close to zero after 5 - 10 steps of backpropagation. This makes it hard for simple recurrent neural networks to learn any long term patterns.
Many architectures were proposed to deal with the vanishing gradients. Among those, the long short term memory (LSTM) recurrent neural network (Hochreiter & Schmidhuber, 1997) is a modified version of simple recurrent network which has obtained promising results on hand writing recognition (Graves & Schmidhuber, 2009) and phoneme classification (Graves & Schmidhuber, 2005). LSTM relies on a fairly sophisticated structure made of gates which control flow of information to hidden neurons. This allows the network to potentially remember information for longer periods. Another interesting direction which was considered is to exploit the structure of the Hessian matrix with respect to the parameters to avoid vanishing gradients. This can be achieved by using second-order methods designed for non-convex objective functions (see section 7 in LeCun et al. (1998)). Unfortunately, there is no clear theoretical justification why using the Hessian matrix would help, nor there is, to the best of our knowledge, any conclusive thorough empirical study on this topic.
In this paper, we propose a simple modification of the SRN to partially solve the vanishing gradient problem. In Section 2, we demonstrate that by simply constraining a part of the recurrent matrix to be close to identity, we can drive some hidden units, called context units to behave like a cache model which can capture long term information similar to the topic of a text (Kuhn & De Mori, 1990). In Section 3, we show that our model can obtain competitive performance compared to the state-of-the-art sequence prediction model, LSTM, on language modeling datasets.
2.1 Simple recurrent network
We consider sequential data that comes in the form of discrete tokens, such as characters or words. We assume a fixed dictionary containing tokens. Our goal is to design a model which is able to predict the next token in the sequence given its past. In this section, we describe the simple recurrent network (SRN) model popularized by Elman (1990), and which is the cornerstone of this work.
A SRN consists of an input layer, a hidden layer with a recurrent connection and an output layer (see Figure 1-a). The recurrent connection allows the propagation through time of information about the state of the hidden layer. Given a sequence of tokens, a SRN takes as input the one-hot encoding of the current token and predicts the probability of next one. Between the current token representation and the prediction, there is a hidden layer with units which store additional information about the previous tokens seen in the sequence. More precisely, at each time , the state of the hidden layer is updated based on its previous state and the encoding of the current token, according to the following equation:
where is the sigmoid function applied coordinate wise, is the token embedding matrix and is the matrix of recurrent weights. Given the state of these hidden units, the network then outputs the probability vector of the next token, according to the following equation:
where is the soft-max function and is the output matrix. In some cases, the size of the dictionary can be significant (e.g., more than 100K tokens for standard language modeling tasks) and computing the normalization term of the soft-max function is often the bottle-neck of this type of architecture. A common trick introduced in Goodman (2001a) is to replace the soft-max function by a hierarchical soft-max. We use a simple hierarchy with two levels, by binning the tokens into clusters with same cumulative word frequency (Mikolov et al., 2011). This reduces the complexity of computing the soft-max from to about , but at the cost of lower performance (around 10 loss in perplexity). We will mention explicitly when we use this approximation in the experiments.
The model is trained by using stochastic gradient descent method with back-propagation through time (Rumelhart et al., 1985; Williams & Zipser, 1995; Werbos, 1988). We use gradient renormalization to avoid gradient explosion. In practice, this strategy is equivalent to gradient clipping since gradient explosions happen very rarely when reasonable hyper-parameters are used. The details of the implementation are given in the experiment section.
It is generally believed that using a strong nonlinearity is necessary to capture complex patterns appearing in real-world data. In particular, the class of mapping that a neural network can learn between the input space and the output space depends directly on these nonlinearities (along with the number of hidden layers and their sizes). However, these nonlinearities also introduce the so-called vanishing gradient problem in recurrent networks. The vanishing gradient problem states that as the gradients get propagated back in time, their magnitude quickly shrinks close to zero. This makes learning longer term patterns difficult, resulting in models which fail to capture the surrounding context. In the next section, we propose a simple extension of SRN to circumvent this problem, yielding a model that can retain information about longer context.
2.2 Context features
In this section, we propose an extension of SRN by adding a hidden layer specifically designed to capture longer term dependencies. We design this layer following two observations: (1) the nonlinearity can cause gradients to vanish, (2) a fully connected hidden layer changes its state completely at every time step.
SRN uses a fully connected recurrent matrix which allows complex patterns to be propagated through time but suffers from the fact that the state of the hidden units changes rapidly at every time step. On the other hand, using a recurrent matrix equal to identity and removing the nonlinearity would keep the state of the hidden layer constant, and every change of the state would have to come from external inputs. This should allow to retain information for longer period of time. More precisely, the rule would be:
where is the context embedding matrix. This solution leads to a model which cannot be trained efficiently. Indeed, the gradient of the recurrent matrix would never vanish, which would require propagation of the gradients up to the beginning of the training set.
Many variations around this type of memory have been studied in the past (see Mozer (1993) for an overview of existing models). Most of these models are based on SRN with no off-diagonal recurrent connections between the hidden units. They differ in how the diagonal weights of the recurrent matrix are constrained. Recently, Pachitariu & Sahani (2013) showed that this type of architecture can achieve performance similar to a full SRN when the size of the dataset and of the model are small. This type of architecture can potentially retain information about longer term statistics, such as the topic of a text, but it does not scale well to larger datasets (Pachitariu & Sahani, 2013). Besides, it can been argued that purely linear SRNs with learned self-recurrent weights will perform very similarly to a combination of cache models with different rates of information decay (Kuhn & De Mori, 1990). Cache models compute probability of the next token given a bag-of-words (unordered) representation of longer history. They are well known to perform strongly on small datasets (Goodman, 2001b). Mikolov & Zweig (2012) show that using such contextual features as additional inputs to the hidden layer leads to a significant improvement in performance over the regular SRN. However in their work, the contextual features are pre-trained using standard NLP techniques and not learned as part of the recurrent model.
In this work, we propose a model which learns the contextual features using stochastic gradient descent. These features are the state of a hidden layer associated with a diagonal recurrent matrix similar to the one presented in Mozer (1993). In other words, our model possesses both a fully connected recurrent matrix to produce a set of quickly changing hidden units, and a diagonal matrix that that encourages the state of the context units to change slowly (see the detailed model in Figure 1-b). The fast layer (called hidden layer in the rest of this paper) can learn representations similar to n-gram models, while the slowly changing layer (called context layer) can learn topic information, similar to cache models. More precisely, denoting by the state of the context units at time , the update rules of the model are:
where is a parameter in and is a matrix. Note that there is no nonlinearity applied to the state of the context units. The contextual hidden units can be seen as an exponentially decaying bag of words representation of the history. This exponential trace memory (as denoted by Mozer (1993)) has been already proposed in the context of simple recurrent networks (Jordan, 1987; Mozer, 1989).
A close idea to our work is to use so-called ”leaky integration” neurons (Jaeger et al., 2007), which also forces the neurons to change their state slowly, however without the structural constraint of SCRN. It was evaluated on the same dataset as we use further (Penn Treebank) by Bengio et al. (2013). Interestingly, the results we observed in our experiments show much bigger gains over stronger baseline using our model, as will be shown later.
Alternative Model Interpretation.
If we consider the context units as additional hidden units (with no activation function), we can see our model as a SRN with a constrained recurrent matrix on both hidden and context units:
where is the identity matrix and is a square matrix of size , i.e., the sum of the number of hidden and context units. This reformulation shows explicitly our structural modification of the Elman SRN (Elman, 1990): we constrain a diagonal block of the recurrent matrix to be equal to a reweighed identity, and keep an off-diagonal block equal to 0. For this reason, we call our model Structurally Constrained Recurrent Network (SCRN).
Adaptive Context Features.
Fixing the weight to be constant in Eq. (4) forces the hidden units to capture information on the same time scale. On the other hand, if we allow this weight to be learned for each unit, we can potentially capture context from different time delays (Pachitariu & Sahani, 2013). More precisely, we denote by the recurrent matrix of the contextual hidden layer, and we consider the following update rule for the state of the contextual hidden layer :
where is a diagonal matrix with diagonal elements in . We suppose that these diagonal elements are obtained by applying a sigmoid transformation to a parameter vector , i.e., diag. This parametrization naturally forces the diagonal weights to stay strictly between 0 and 1.
We study in the following section in what situations does learning of the weights help. Interestingly, we show that learning of the self-recurrent weights does not seem to be important, as long as one uses also the standard hidden layer in the model.
We evaluate our model on the language modeling task for two datasets. The first dataset is the Penn Treebank Corpus, which consists of 930K words in the training set. The pre-processing of data and division to training, validation and test parts are the same as in (Mikolov et al., 2011). The state-of-the-art performance on this dataset has been achieved by Zaremba et al. (2014), using combination of many big, regularized LSTM recurrent neural network language models. The LSTM networks were first introduced to language modeling by Sundermeyer et al. (2012).
The second dataset, which is moderately sized, is called Text8. It is composed of a pre-processed version of the first 100 million characters from Wikipedia dump. We did split it into training part (first 99M characters) and development set (last 1M characters) that we use to report performance. After that, we constructed the vocabulary and replaced all words that occur less than 5 times by UNK token. The resulting vocabulary size is about 44K. To simplify reproducibility of our results, we released both the SCRN code and the scripts which construct the datasets 111The SCRN code can be downloaded at http://github.com/facebook/SCRNNs.
In this section we compare the performance of our proposed model against standard SRNs, and LSTM RNNs which are becoming the architecture of choice for modeling sequential data with long-term dependencies.
3.1 Implementation Details.
We used Torch library and implemented our proposed model following the graph given in Figure 1-b. Note that following the alternative interpretation of our model with the recurrent matrix defined in Eq. 8, our model could be simply implemented by modifying a standard SRN. We fix at 0.95 unless stated otherwise. The number of backpropagation through time (BPTT) steps is set to 50 for our model and was chosen by parameter search on the validation set. For normal SRN, we use just 10 BPTT steps because the gradients vanish faster. We do a stochastic gradient descent after every 5 forward steps. Our model is trained with a batch gradient descent of size 32, and a learning rate of 0.05. We divide the learning rate by 1.5 after each training epoch when the validation error does not decrease.
3.2 Results on Penn Treebank Corpus.
We first report results on the Penn Treebank Corpus using both small and moderately sized models (with respect to the number of hidden units). Table 1 shows that our structurally constrained recurrent network (SCRN) model can achieve performance comparable with LSTM models on small datasets with relatively small numbers of parameters. It should be noted that the LSTM models have significantly more parameters for the same size of hidden layer, making the comparison somewhat unfair - with the input, forget and output gates, the LSTM has about 4x more parameters than SRN with the same size of hidden layer.
Comparison to ”leaky neurons” is also in favor of SCRN: Bengio et al. (2013) report perplexity reduction from 136 (SRN) to 129 (SRN + leaky neurons), while for the same dataset, we observed much bigger improvement, going from perplexity 129 (SRN) down to 115 (SCRN).
Table 1 also shows that SCRN outperforms the SRN architecture even with much less parameters. This can be seen by comparing performance of SCRN with 40 hidden and 10 contextual units (test perplexity 127) versus SRN with 300 hidden units (perplexity 129). This suggests that imposing a structure on the recurrent matrix allows the learning algorithm to capture additional information. To obtain further evidence that this additional information is of a longer term character, we did further run experiments on the Text8 dataset that contains various topics, and thus the longer term information affects the performance on this dataset much more.
|Model||hidden||context||Validation Perplexity||Test Perplexity|
|Ngram + cache||-||-||-||125|
3.2.1 Learning Self-Recurrent Weights.
We evaluate influence of learning the diagonal weights of the recurrent matrix for the contextual layer. For the following experiments, we used a hierarchical soft-max with 100 frequency-based classes on the Penn Treebank Corpus to speedup the experiments. In Table 2, we show that when the size of the hidden layer is small, learning the diagonal weights is crucial. This result confirms the findings in Pachitariu & Sahani (2013). However, as we increase the size of our model and use sufficient number of hidden units, learning of the self-recurrent weights does not give any significant improvement. This indicates that learning the weights of the contextual units allows these units to be used as multi-scale representation of the history, i.e., some contextual units can specialize on the very recent history (for example, for close to , the contextual units would be part of a simple bigram language model). With various learned self-recurrent weights, the model can be seen as a combination of cache and bigram models. When the number of standard hidden units is enough to capture short term patterns, learning the self-recurrent weights does not seem crucial anymore.
Keeping this observation in mind we fixed the diagonal weights when working with the Text8 corpus.
|Model||hidden||context||Fixed weights||Adaptive weights|
3.3 Results on Text8.
Our next experiment involves the Text8 corpus which is significantly larger than the Penn Treebank. As this dataset contains various articles from Wikipedia, the longer term information (such as current topic) plays bigger role than in the previous experiments. This is illustrated by the gains when cache is added to the baseline 5-gram model: the perplexity drops from 309 to 229 (26% reduction).
We report experiments with a range of model configurations, with different number of hidden units. In Table 3, we show that increasing the capacity of standard SRNs by adding the contextual features results in better performance. For example, when we add 40 contextual units to SRN with 100 hidden units, the perplexity drops from 245 to 189 (23% reduction). Such model is also much better than SRN with 300 hidden units (perplexity 202).
|Model||hidden||context = 0||context = 10||context = 20||context = 40||context = 80|
In Table 4, we see that when the number of hidden units is small, our model is better than LSTM. Despite the LSTM model with 100 hidden units being larger, the SCRN with 100 hidden and 80 contextual features achieves better performance. On the other hand, as the size of the models increase, we see that the best LSTM model is slightly better than the best SCRN (perplexity 156 versus 161). As the perplexity gains for both LSTM and SCRN over SRN are much more significant than in the Penn Treebank experiments, it seems likely that both models actually model the same kind of patterns in language.
|Model||hidden||context||Perplexity on development set|
In this paper, we have shown that learning longer term patterns in real data using recurrent networks is perfectly doable using standard stochastic gradient descent, just by introducing structural constraint on the recurrent weight matrix. The model can then be interpreted as having quickly changing hidden layer that focuses on short term patterns, and slowly updating context layer that retains longer term information.
Empirical comparison of SCRN to Long Short Term Memory (LSTM) recurrent network shows very similar behavior in two language modeling tasks, with similar gains over simple recurrent network when all models are tuned for the best accuracy. Moreover, SCRN shines in cases when the size of models is constrained, and with similar number of parameters it often outperforms LSTM by a large margin. This can be especially useful in cases when the amount of training data is practically unlimited, and even models with thousands of hidden neurons severely underfit the training datasets.
We believe these findings will help researchers to better understand the problem of learning longer term memory in sequential data. Our model greatly simplifies analysis and implementation of recurrent networks that are capable of learning longer term patterns. Further, we published the code that allows to easily reproduce experiments described in this paper.
At the same time, it should be noted that none of the above models is capable of learning truly long term memory, which has a different nature. For example, if we would want to build a model that can store arbitrarily long sequences of symbols and reproduce these later, it would become obvious that this is not doable with models that have finite capacity. A possible solution is to use the recurrent net as a controller of an external memory which has unlimited capacity. For example in (Joulin & Mikolov, 2015), a stack-based memory is used for such task. However, a lot of research needs to be done in this direction before we will develop models that can successfully learn to grow in complexity and size when solving increasingly more difficult tasks.
- Bengio et al. (1994) Bengio, Yoshua, Simard, Patrice, and Frasconi, Paolo. Learning long-term dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994.
- Bengio et al. (2013) Bengio, Yoshua, Boulanger-Lewandowski, Nicolas, and Pascanu, Razvan. Advances in optimizing recurrent networks. In ICASSP, 2013.
- Dahl et al. (2012) Dahl, George E, Yu, Dong, Deng, Li, and Acero, Alex. Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition. Audio, Speech, and Language Processing, IEEE Transactions on, 20(1):30–42, 2012.
- Elman (1990) Elman, Jeffrey L. Finding structure in time. Cognitive science, 14(2):179–211, 1990.
- Goodman (2001a) Goodman, Joshua. Classes for fast maximum entropy training. In Acoustics, Speech, and Signal Processing, 2001. Proceedings.(ICASSP’01). 2001 IEEE International Conference on, volume 1, pp. 561–564. IEEE, 2001a.
- Goodman (2001b) Goodman, Joshua T. A bit of progress in language modeling. Computer Speech & Language, 15(4):403–434, 2001b.
- Graves & Schmidhuber (2009) Graves, Alex and Schmidhuber, Juergen. Offline handwriting recognition with multidimensional recurrent neural networks. In Advances in Neural Information Processing Systems, pp. 545–552, 2009.
- Graves & Schmidhuber (2005) Graves, Alex and Schmidhuber, Jürgen. Framewise phoneme classification with bidirectional lstm and other neural network architectures. Neural Networks, 18(5):602–610, 2005.
- Hochreiter (1998) Hochreiter, Sepp. The vanishing gradient problem during learning recurrent neural nets and problem solutions. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 6(02):107–116, 1998.
- Hochreiter & Schmidhuber (1997) Hochreiter, Sepp and Schmidhuber, Jürgen. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.
- Jaeger et al. (2007) Jaeger, Herbert, Lukoševičius, Mantas, Popovici, Dan, and Siewert, Udo. Optimization and applications of echo state networks with leaky-integrator neurons. Neural Networks, 20(3):335–352, 2007.
- Jordan (1987) Jordan, Michael I. Attractor dynamics and parallelism in a connectionist sequential machine. Proceedings of the Eighth Annual Conference of the Cognitive Science Society, pp. 531–546, 1987.
- Joulin & Mikolov (2015) Joulin, Armand and Mikolov, Tomas. Inferring algorithmic patterns with stack-augmented recurrent nets. arXiv preprint arXiv:1503.01007, 2015.
- Koehn et al. (2007) Koehn, Philipp, Hoang, Hieu, Birch, Alexandra, Callison-Burch, Chris, Federico, Marcello, Bertoldi, Nicola, Cowan, Brooke, Shen, Wade, Moran, Christine, Zens, Richard, et al. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the ACL on Interactive Poster and Demonstration Sessions, pp. 177–180. Association for Computational Linguistics, 2007.
- Kuhn & De Mori (1990) Kuhn, Roland and De Mori, Renato. A cache-based natural language model for speech recognition. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 12(6):570–583, 1990.
- LeCun et al. (1998) LeCun, Yann, Bottou, Leon, Orr, Genevieve, and Müller, Klaus. Efficient backprop. Neural Networks: Tricks of the Trade, pp. 546–546, 1998.
- Mikolov (2012) Mikolov, Tomáš. Statistical language models based on neural networks. PhD thesis, Ph. D. thesis, Brno University of Technology, 2012.
- Mikolov & Zweig (2012) Mikolov, Tomas and Zweig, Geoffrey. Context dependent recurrent neural network language model. In SLT, pp. 234–239, 2012.
- Mikolov et al. (2011) Mikolov, Tomas, Kombrink, Stefan, Burget, Lukas, Cernocky, JH, and Khudanpur, Sanjeev. Extensions of recurrent neural network language model. In Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on, pp. 5528–5531. IEEE, 2011.
- Mozer (1989) Mozer, Michael C. A focused back-propagation algorithm for temporal pattern recognition. Complex systems, 3(4):349–381, 1989.
- Mozer (1993) Mozer, Michael C. Neural net architectures for temporal sequence processing. In Santa Fe Institute Studies in The Sciences of Complexity, volume 15, pp. 243–243. Addison-Wessley Publishing CO, 1993.
- Nair & Hinton (2010) Nair, Vinod and Hinton, Geoffrey E. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pp. 807–814, 2010.
- Pachitariu & Sahani (2013) Pachitariu, Marius and Sahani, Maneesh. Regularization and nonlinearities for neural language models: when are they needed? arXiv preprint arXiv:1301.5650, 2013.
- Rumelhart et al. (1985) Rumelhart, David E, Hinton, Geoffrey E, and Williams, Ronald J. Learning internal representations by error propagation. Technical report, DTIC Document, 1985.
- Simonyan & Zisserman (2014) Simonyan, Karen and Zisserman, Andrew. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems, pp. 568–576, 2014.
- Sundermeyer et al. (2012) Sundermeyer, Martin, Schlüter, Ralf, and Ney, Hermann. Lstm neural networks for language modeling. In INTERSPEECH, 2012.
- Werbos (1988) Werbos, Paul J. Generalization of backpropagation with application to a recurrent gas market model. Neural Networks, 1(4):339–356, 1988.
- Williams & Zipser (1995) Williams, Ronald J and Zipser, David. Gradient-based learning algorithms for recurrent networks and their computational complexity. Back-propagation: Theory, architectures and applications, pp. 433–486, 1995.
- Young et al. (1997) Young, Steve, Evermann, Gunnar, Gales, Mark, Hain, Thomas, Kershaw, Dan, Liu, Xunying, Moore, Gareth, Odell, Julian, Ollason, Dave, Povey, Dan, et al. The HTK book, volume 2. Entropic Cambridge Research Laboratory Cambridge, 1997.
- Zaremba et al. (2014) Zaremba, Wojciech, Sutskever, Ilya, and Vinyals, Oriol. Recurrent neural network regularization. arXiv preprint arXiv:1409.2329, 2014.