Rotational Unit of Memory: A Novel Representation Unit for RNNs with Scalable Applications

An important element of the ongoing neural revolution in Natural Language Processing (NLP) is the rise of Recurrent Neural Networks (RNNs), which have become a standard tool for addressing a number of tasks ranging from language modeling, part-of-speech tagging and named entity recognition to neural machine translation, text summarization, question answering, and building chatbots/ dialog systems.

As standard RNNs suffer from exploding/ vanishing gradient problems, alternatives such as long short-term memory (LSTM) (Hochreiter and Schmidhuber, 1997) or gated recurrent units (GRUs) (Cho et al., 2014) have been proposed and have now become standard.

Nevertheless, LSTMs and GRUs fail to demonstrate really long-term memory capabilities or efficient recall on synthetic tasks (see Figure 1). Figure 1 shows that when RNN units are fed a long string (e.g., emojis in Figure 1(a)), they struggle to represent the input in their memory, which results in recall or copy mistakes. The origins of these issues are two-fold: (i) a single hidden state cannot memorize complicated sequential dynamics and (ii) the hidden state is not manipulated well, resulting in information loss. Typically, these are addressed separately: by using external memory for (i), and gated mechanisms for (ii).

Here, we solve (i) and (ii) jointly by proposing a novel RNN unit, Rotational Unit of Memory (RUM), that manipulates the hidden state by rotating it in an Euclidean space, resulting in a better information flow. This remedy to (ii) affects (i) to the extent that the external memory is less needed. As a proof of concept, in Figure 1(a), RUM recalls correctly a faraway emoji.

We further show that RUM is fit for real-world NLP tasks. In Figure 1(b), a RUM-based seq2seq model produces a better summary than what a standard LSTM-based seq2seq model yields. In this particular example, LSTM falls into the well-known trap of repeating information close to the end, whereas RUM avoids it. Thus, RUM can be seen as a more “well-rounded” alternative to LSTM.

Given the example from Figure 1, we ask the following questions: Does the long-term memory’s advantage for synthetic tasks such as copying and recall translate to improvements for real-world NLP problems? Can RUM solve issues (i) and (ii) more efficiently? Does a theoretical advance improve real-world applications?

We propose RUM as the answer to these questions via experimental observations and mathematical intuition. We combine concepts from unitary learning and associative memory to utilize the theoretical advantages of rotations, and then we show promising applications to hard NLP tasks. Our evaluation of RUM is organized around a sequence of tests: (1) Pass a synthetic memory copying test; (2) Pass a synthetic associative recall test; (3) Show promising performance for question answering on the bAbI data set; (4) Improve the state-of-the-art for character-level language modeling on the Penn Treebank; (5) Perform effective seq2seq text summarization, training on the difficult CNN / Daily Mail summarization corpus.

To the best of our knowledge, there is no previous work on RNN units that shows such promising performance, both theoretical and practical. Our contributions can be summarized as follows: (i) We propose a novel representation unit for RNNs based on an idea not previously explored in this context—rotations. (ii) We show theoretically and experimentally that our unit models much longer distance dependencies than LSTM and GRU, and can thus solve tasks such as memory recall and memory copying much better. (iii) We further demonstrate that RUM can be used as a replacement for LSTM/GRU in real-world problems such as language modeling, question answering, and text summarization, yielding results comparable to the state of the art.¹

Our work rethinks the concept of gated models. LSTM and GRU are the most popular such models, and they learn to generate gates—such as input, reset, and update gates—that modify the hidden state through element-wise multiplication and addition. We manipulate the hidden state in a completely different way: Instead of gates, we learn directions in the hidden space towards which we rotate it.

Moreover, because rotations are orthogonal, RUM is implicitly orthogonal, meaning that RUM does not parametrize the orthogonal operation, but rather extracts it from its own components. Thus, RUM is also different from explicitly orthogonal models such as uRNN, EURNN, GORU, and all other RNN units that parametrize their norm-preserving operation (see below).

Rotations have fundamental applications in mathematics (Artin, 2011; Hall, 2015) and physics (Sakurai and Napolitano, 2010). In computer vision, rotational matrices and quaternions contain valuable information and have been used to estimate object poses (Katz, 2001; Shapiro and Stockman, 2001; Kuipers, 2002). Recently, efficient, accurate and rotationally invariant representation units have been designed for convolutional neural networks (Worrall et al., 2017; Cohen et al., 2018; Weiler et al., 2018). Unlike that work, we use rotations to design a new RNN unit with application to NLP, rather than vision.

Unitary learning approaches the problem of vanishing and exploding gradients, which obstruct learning of really long-term dependencies (Bengio et al., 1994). Theoretically, using unitary and orthogonal matrices will keep the norm of the gradient: the absolute value of their eigenvalues is raised to a high power in the gradient calculation, but it equals one, so it neither vanishes, nor explodes. Arjovsky et al. (2016) (unitary RNN, or uRNN) and Jing et al. (2017b) (efficient unitary RNN, or EURNN) used parameterizations to form the unitary spaces. Wisdom et al. (2016) applied gradient projection onto a unitary manifold. Vorontsov et al. (2017) used penalty terms as a regularization to restrict the matrices to be unitary. Unfortunately, such theoretical approaches struggle to perform outside of the domain of synthetic tasks, and fail at simple real- world tasks such as character-level language modeling (Jing et al., 2017a). To alleviate this issue, Jing et al. (2017a) combined a unitary parametrization with gates, thus yielding a gated orthogonal recurrent unit (GORU), which provides a ‘‘forgetting mechanism,’’ required by NLP tasks.

Among pre-existing RNN units, RUM is most similar to GORU in spirit because both models transform (significantly) GRU. Note, however, that GORU parametrizes an orthogonal operation whereas RUM extracts an orthogonal operation in the form of a rotation. In this sense, to parallel our model’s implicit orthogonality to the literature, RUM is a ‘‘firmware’’ structure rather than a ‘‘learnware’’ structure, as discussed in (Balduzzi and Ghifary, 2016).

Associative memory modeling provides a large variety of input encodings in a neural network for effective pattern recognition (Kohonen, 1974; Krotov and Hopfield, 2016). It is particularly appealing for RNNs because their memory is in short supply. RNNs often circumvent this by using external memory in the form of an attention mechanism (Bahdanau et al., 2015; Hermann et al., 2015). Another alternative is the use of neural Turing machines (Graves et al., 2014, 2016). In either case, this yields an increase in the size of the model and makes training harder.

Recent advances in associative memory (Plate, 2003; Danihelka et al., 2016; Ba et al., 2016a; Zhang and Zhou, 2017) suggest that its updates can be learned efficiently with backpropagation through time (Rumelhart et al., 1986). For example, Zhang and Zhou (2017) used weights that are dynamically updated by the input sequence. By treating the RNN weights as memory determined by the current input data, a larger memory size is provided and fewer trainable parameters are required.

Note that none of these methods used rotations to create the associative memory. The novelty of our model is that it exploits the simple and fundamental multiplicative closure of rotations to generate rotational associative memory for RNNs. As a result, an RNN that uses our RUM units yields state-of-the-art performance for synthetic associative recall while using very few parameters.

Successful RNNs require well-engineered manipulations of the hidden state h_t at time step t. We approach this mathematically, considering h_t as a real vector in an N_h-dimensional Euclidean space, where N_h is the dimension of the ‘‘hidden’’ state $RNh$⁠. Note that there is an angle between two vectors in $RNh$ (the cosine of that angle can be calculated as a normalized dot product ‘‘⋅’’). Moreover, we can associate a unique angle to h_t with respect to some reference vector. Thus, a hidden state can be characterized by a magnitude, i.e., L₂-norm ‘‘∥.∥’’, and a phase, i.e., angle with respect to the reference vector. Thus, if we devise a mechanism to generate reference vectors at any given time step, we would enable rotating the hidden state with respect to the generated reference.

This rethinking of $RNh$ allows us to propose a powerful learning representation: Instead of following the standard way of learning to modify the norm of h_t through multiplication by gates and self-looping (as in LSTM), we learn to rotate the hidden state, thereby changing the phase, but preserving the magnitude. The benefits of using such phase-learning representation are two-fold: (i) the preserved magnitude yields stable gradients, which in turn enables really long-term memory, and (ii) there is always a sequence of rotations that can bring the current phase to a desired target one, thus enabling effective recall of information.

In order to achieve this, we need a phase-learning transformation that is also differentiable. A simple and efficient approach is to compute the angle between two special vectors, and then to update the phase of the hidden state by rotating it with the computed angle.

We let the RNN generate the special vectors at time step t (i) by linearly embedding the RNN input $xt∈RNx$ to an embedded input$ε~t∈RNh$⁠, and (ii) by obtaining a target memoryτ_t as a linear combination of the current input x_t (projected in the hidden space) and the previous history h_t−1 (after a linear transformation).

We propose a function $Rotation:RNh×RNh→RNh×Nh,$ which implements this idea. Rotation takes a pair of column vectors (a, b) and returns the rotational matrix R from a to b. If a and b have the same orientation, then R is the identity matrix; otherwise, the two vectors form a plane span(a, b). Our operation projects and rotates in that plane, leaving everything else intact, as shown in Figure 2(a) for $a=ε~$ and b = τ (for simplicity, we drop the time indices).

We make the associative memory from Equation (2) tunable through the parameter λ ∈{0, 1}, which serves to switch the associative memory off and on. To the best of our knowledge, our model is the first RNN to explore such multiplicative associative memory.

Note that even though R_t acts as an additional memory state, there are no additional parameters in RUM: The parameters are only within the Rotation operation. As the same Rotation appears at each recursive step (2), the parameters are shared.

Figure 2(b) shows a sketch of the connections in the RUM cell. RUM consists of an update gate $u∈RNh$ that has the same function as in GRU. However, instead of a reset gate, the model learns the memory target $τ∈RNh$⁠. RUM also learns to embed the input vector $x∈RNx$ into $RNh$ to yield $ε~∈RNh.$ Hence, Rotation encodes the rotation between the embedded input and the target, which is accumulated in the associative memory unit $Rt∈RNh×Nh$ (originally initialized to the identity matrix). Here, λ is a non-negative integer that is a hyper-parameter of the model. The orthogonal matrix R_t acts on the state h to produce an evolved hidden state $h~.$ Finally, RUM calculates the new hidden state via u, just as in GRU. The RUM equations are given in Algorithm 1. The orthogonal matrix $R(εt~,τ)$ conceptually takes the place of a weight kernel acting on the hidden state in GRU.

For the sake of clarity, let us consider a simplified version of RUM, where W ≡ R_k is a rotation matrix, and let us use spectral norm for matrices. By orthogonality, we have ∥W^†∥ = 1. Then, the norm of the update ∥∂C/∂h_t∥ is bounded by $∥∂C/∂hT∥∥W†∥T−t∏k=1T−1∥D(k)∥$⁠, which simplifies to $∥∂C/∂hT∥∏k=1T−1∥D(k)∥$⁠. Hence, if the norm of ∥D^(k)∥ is strictly less than one, we would witness vanishing gradients (for large T), which we aim to avoid by choosing a proper activation.

Hence, we compare four well-known activations f: ReLU, $tanh$⁠, sigmoid, and softsign. Figure 3 shows their derivatives. As long as some value is positive, the ReLU derivative will be one, and thus ∥D^(k)∥ = 1. This means that ReLU is potentially a good choice. Because RUM is closer to GRU, which makes the analysis more complicated, we conduct ablation studies on non-linear activations and on the importance of the update gate throughout Section 4.

We call it time normalization, as we usually feed mini-batches to the RNN during learning that have the shape (N_b, N_T), where N_b is the size of the batch and N_T is the length of the sequence. Empirically, fixing η to be a small number stabilizes training, and we found that values centered around η = 1.0 work well. This is an optional component in RUM, as typically $ht′$ does not blow up. In our experiments, we only needed it for our character-level language modeling, which mixes RUM and LSTM units.

We now describe two kinds of experiments based (i) on synthetic and (ii) on real-world tasks. The former test the representational power of RUMs vs. LSTMs/GRUs, and the latter test whether RUMs also perform well for real-world NLP problems.

Copying memory task (A) is a standard testbed for the RNN’s capability for long-term memory (Hochreiter and Schmidhuber, 1997; Arjovsky et al., 2016; Henaff et al., 2016). Here, we follow the experimental set-up in Jing et al. (2017b).

Data. The alphabet of the input consists of symbols {a_i}, i ∈{0, 1, ⋯ ,n − 1, n, n + 1}, the first n of which represent data for copying, and the remaining two forming ‘‘blank’’ and ‘‘marker’’ symbols, respectively. In our experiments, we set n = 8 and the data for copying is the first 10 symbols of the input. The RNN model is expected to output ‘‘blank’’ during T = 500 delay steps and, after the ‘‘marker’’ appears in the input, to output (copy) sequentially the first 10 input symbols. The train/test split is 50,000/500 examples.

Models. We test RNNs built using various types of units: LSTM (Hochreiter and Schmidhuber, 1997), GRU (Cho et al., 2014), uRNN (Wisdom et al., 2016), EURNN (Jing et al., 2017b), GORU (Jing et al., 2017a), and RUM (ours) with λ ∈{0,1} and η ∈{1.0,N/A}. We train with a batch size of 128 and an RMSProp with a 0.9 decay rate, and we try learning rates from {0.01, 0.001, 0.0001}. We found that LSTM and GRU fail for all learning rates, EURNN is unstable for large learning rates, and RUM is stable for all learning rates. Thus, we use 0.001 for all units except for EURNN, for which we use 0.0001.

Results. In Figure 4, we show the cross-entropy loss for delay time T = 500. Note that LSTM and GRU hit a predictable baseline of memoryless strategy, equivalent to random guessing.⁴ We can see that RUM improves over the baseline and converges to 100% accuracy. For the explicit unitary models, EURNN and uRNN also solve the problem in just a few steps, and GORU converges slightly faster than RUM.

Next, we study why RUM units can solve the task, whereas LSTM/GRU units cannot. In Figure 4, we also test a RUM model (called RUM′) without a flexible target memory and embedded input, that is, the weight kernels that produce τ_t and $ε~t$ are kept constant. We observe that the model does not learn well (converges extremely slowly). This means that learning to rotate the hidden state by having control over the angles used for rotations is indeed needed.

Controlling the norm of the hidden state is also important. The activations of LSTM and GRU are sigmoid and tanh, respectively, and both are bounded. RUM uses ReLU, which allows larger hidden states (nevertheless, note that RUM with the bounded tanh also yields 100% accuracy). We observe that, when we remove the normalization, RUM converges faster compared with using η = 1.0. Having no time normalization means larger spikes in the cross-entropy and increased risk of exploding loss. EURNN and uRNN are exposed to this, while RUM uses a tunable reduction of the risk through time normalization.

We also observe the benefits of tuning the associative rotational memory. Indeed, a RUM with λ = 1 has a smaller hidden size, N_h = 100, but it learns much faster than a RUM with λ = 0. It is possible that the accumulation of phase via λ = 1 enables faster really long-term memory.

Finally, we would like to note that removing the update gate or using tanh and softsign activations do not hurt performance.

Associative recall task (B) is another testbed for long-term memory. We follow the settings in Ba et al. (2016a) and Zhang and Zhou (2017).

Data. The sequences for training are random, and consist of pairs of letters and digits. We set the query key to always be a letter. We fix the size of the letter set to half the length of the sequence, the digits are from 0 to 9. No letter is repeated. In particular, the RNN is fed a sequence of letter–digit pairs followed by the separation indicator ‘‘??’’ and a query letter (key), e.g., ‘‘a1s2d3f4g5??d’’. The RNN is supposed to output the digit that follows the query key (‘‘d’’ in this example): It needs to find the query key and then to output the digit that follows (‘‘3’’ in this example). The train/dev/test split is 100k/10k/20k examples.

Models. We test LSTM, GRU, GORU, FW-LN (Ba et al., 2016a), WeiNet (Zhang and Zhou, 2017), and RUM (λ = 1, η = 0). All the models have the same hidden state N_h = 50 for different lengths T. We train for 100k epochs with a batch size of 128, RMSProp as an optimizer, and a learning rate of 0.001 (selected using hyper-parameter search).

Results.Table 1 shows the results. We can see that LSTM and GRU are unable to recall the digit correctly. Even GORU, which learns the copying task, fails to solve the problem. FW-LN, WeiNet, and RUM can learn the task for T = 30. For RUM, it is necessary that λ = 1, as for λ = 0 its performance is similar to that of LSTM and GORU. WeiNet and RUM are the only known models that can learn the task for the challenging 50 input characters. Note that RUM yields 100% accuracy with 40% fewer parameters.

The benefit of the associative memory is apparent from the temperature map in Figure 5(a), where we can see that the weight kernel for the target memory has a clear diagonal activation. This suggests that the model learns how to rotate the hidden state in the Euclidean space by observing the sequence encoded in the hidden state. Note that none of our baseline models exhibit such a pattern for the weight kernels.

Figure 5(b) shows the evolution of the rotational behavior during the 53 time steps for a model that does not learn the task. We can see that $cosθ$ is small and biased towards 0.2. Figure 5(c) shows the evolution of a model with associative memory (λ = 1) that does learn the task. Note that these distributions have a wider range that is more uniform.

Also, there are one or two $cosθ$ instances close to 1.0 per distribution, that is, the angle is close to zero and the hidden state is rotated only marginally. The distributions in Figure 5(c) yield more varied representations.

Question answering (C) is typically done using neural networks with external memory, but here we use a vanilla RNN with and without attention.

Data. We use the bAbI Question Answering data set (Weston et al., 2016), which consists of 20 subtasks, with 9k/1k/1k examples for train/ dev/test per subtask. We train a separate model for each subtask. We tokenize the text (at the word and at the sentence level), and then we concatenate the story and the question.

For the word level, we embed the words into dense vectors, and we feed them into the RNN. Hence, the input sequence can be labeled as ${x1(s),…,xn(s),x1(q),…,xm(q)}$⁠, where the story has n words and the question has m words.

For the sentence level, we generate sentence embeddings by averaging word embeddings. Thus, the input sequence for a story with n sentences is ${x1(s),…,xn(s),x(q)}$⁠.

Attention mechanism for sentence level. We use simple dot-product attention (Luong et al., 2015): ${pt}0≤t≤n:=softmax({h(q)⋅ht(s)}0≤t≤n)$⁠. The context vector $c:=∑t=0nptht(s)$ is then passed, together with the query vector, to a dense layer.

Models. We compare uRNN, EURNN, LSTM, GRU, GORU, and RUM (with η = N/A in all experiments). The RNN model outputs the prediction at the end of the question through a softmax layer. We use a batch size of 32 for all 20 subtasks. We train the model using Adam optimizer with a learning rate of 0.001 (Kingma and Ba, 2015). All embeddings (word and sentence) are 64-dimensional. For each subset, we train until convergence on the dev set, without other regularization. For testing, we report the average accuracy over the 20 subtasks.

Results.Table 2 shows the average accuracy on the 20 bAbI tasks. Without attention, RUM outperforms LSTM/GRU and all unitary baseline models by a sizable margin both at the word and at the sentence level. Moreover, RUM without attention (line 14) outperforms all models except for attnLSTM. Furthermore, LSTM and GRU benefit the most from adding attention (lines 10–11), while the phase-coded models (lines 9, 12–15) obtain only a small boost in performance or even a decrease (e.g., in line 13). Although RUM (line 14) shares the best accuracy with LSTM (line 10), we hypothesize that a ‘‘phase-inspired’’ attention might further boost RUM’s performance.⁵

Language modeling [character-level] (D) is an important testbed for RNNs (Graves, 2013).

Data. The Penn Treebank (PTB) corpus is a collection of articles from The Wall Street Journal (Marcus et al., 1993), with a vocabulary of 10k words (using 50 different characters). We use a train/dev/test split of 5.1M/400k/450k tokens, and we replace rare words with <unk>. We feed 150 tokens at a time, and we use a batch size of 128.

Models. We incorporate RUM into a recent high-level model: Fast-Slow RNN (FS-RNN) (Mujika et al., 2017). The FS-RNN-k architecture consists of two hierarchical layers: one of them is a ‘‘fast’’ layer that connects k RNN cells F₁, …, F_k in series; the other is a ‘‘slow’’ layer that consists of a single RNN cell S. The organization is roughly as follows: F₁ receives the input from the mini-batch and feeds its state into S, S feeds its state into F₂, and so on; finally, the output of F_k is a probability distribution over characters. FS-RUM-2 uses fast cells (all LSTM) with hidden size of 700 and a slow cell (RUM) with a hidden state of size 1000, time normalization η = 1.0, and λ = 0. We also tried to use associative memory λ = 1 or to avoid time normalization, but we encountered exploding loss at early training stages. We optimized all hyper-parameters on the dev set.

Additionally, we tested FS-EURNN-2, i.e., the slow cell is EURNN with a hidden size of 2000, and FS-GORU-2 with a slow cell GORU with a hidden size of 800 (everything else remains as for FS-RUM-2). As the learning phases are periodic, there is no easy regularization for FS-EURNN-2 or FS-GORU-2.

For FS-RNN, we use the hyper-parameter values suggested in Mujika et al. (2017). We further use layer normalization (Ba et al., 2016b) on all states, on the LSTM gates, on the RUM update gate, and on the target memory. We also apply zoneout (Krueger et al., 2017) to the recurrent connections, as well as dropout (Srivastava et al., 2014). We embed each character into a 128-dimensional space (without pre-training).

For training, we use the Adam optimizer with a learning rate of 0.002, we decay the learning rate for the last few training epochs, and we apply gradient clipping with a maximal norm of the gradients equal to 1.0. Finally, we pass the output through a softmax layer.

For testing, we report bits-per-character (BPC) loss on the test dataset, which is the cross-entropy loss but with a binary logarithm.

Our best FS-RUM-2 uses decaying learning rate: 180 epochs with a learning rate of 0.002, then 60 epochs with 0.0001, and finally 120 epochs with 0.00001.

We also test a RUM with η = 1.0, and a two-layer RUM with η = 0.3. The cell zoneout/hidden zoneout/dropout probability is 0.5/0.9/0.35 for FS-RUM-2, and 0.5/0.1/0.65 for the vanilla versions. We train for 100 epochs with a 0.002 learning rate. These values were suggested by Mujika et al. (2017), who used LSTM cells.

Results. In Table 3, we report the BPC loss for character-level language modeling on PTB. For the test split, FS-RUM-2 reduces the BPC for Fast-Slow models by 0.001 points absolute. Moreover, we achieved a decrease of 0.002 BPC points for the validation split using an FS-RUM-2 model with a hidden size of 800 for the slow cell (RUM) and a hidden size of 1100 for the fast cells (LSTM). Our results support a conjecture from the conclusions of Mujika et al. (2017), which states that models with long-term memory, when used as the slow cell, may enhance performance.

Text summarization (E) is the task of reducing long pieces of text to short summaries without losing much information. It is one of the most challenging tasks in NLP (Nenkova and McKeown, 2011), with a number of applications ranging from question answering to journalism (Tatalović, 2018). Text summarization can be abstractive (Nallapati et al., 2016), extractive (Nallapati et al., 2017), or hybrid (See et al., 2017). Advances in encoder-decoder/seq2seq models (Cho et al., 2014; Sutskever et al., 2014) established models based on RNNs as powerful tools for text summarization. Having accumulated knowledge from the ablation and the preparation tasks, we test RUM on this hard real-world NLP task.

Data. We follow the set-up in See et al. (2017) and we use the CNN/ Daily Mail corpus (Hermann et al., 2015; Nallapati et al., 2016), which consist of news stories with reference summaries. On average, there are 781 tokens per story and 56 tokens per summary. The train/dev/test datasets contain 287,226/13,368/11,490 text–summary pairs.

We further experimented with a new data set, which we crawled from the Science Daily Web site, iterating certain patterns of date/time. We successfully extracted 60,900 Web pages, each containing a public story about a recent scientific paper. We extracted the main content, a short summary, and a title from the HTML page using Beautiful Soup. The input story lengthis 488.42 ± 219.47, the target summary length is 45.21 ± 18.60, and the title length is 9.35 ± 2.84. In our experiments, we set the vocabulary size to 50k.

We defined four tasks on this data: (i) s2s, story to summary, (ii) sh2s, shuffled story to summary (we put the paragraphs in the story in a random order); (iii) s2t, story to title; and (iv) oods2s, out-of-domain testing for s2s (i.e., training on CNN / Daily Mail and testing on Science Daily).

Models. We use a pointer-generator network (See et al., 2017), which is a combination of a seq2seq model (Nallapati et al., 2016) with attention (Bahdanau et al., 2015) and a pointer network (Vinyals et al., 2015). We believe that the pointer-generator network architecture to be a good testbed for experiments with a new RNN unit because it enables both abstractive and extractive summarization.

We adopt the model from See et al. (2017) as our LEAD baseline. This model uses a bi-directional LSTM encoder (400 steps) with attention distribution and an LSTM decoder (100 steps for training and 120 steps for testing), with all hidden states being 256-dimensional, and 128-dimensional word embeddings trained from scratch during training. For training, we use the cross-entropy loss for the seq2seq model. For evaluation, we use ROUGE (Lin and Hovy, 2003). We also allow the coverage mechanism proposed in the original paper, which penalizes repetitions and improves the quality of the summaries (marked as ‘‘cov.’’ in Table 4). Following the original paper, we train LEAD for 270k iterations and we turn on the coverage for about 3k iterations at the end to get LEAD cov. We use an Adagrad optimizer with a learning rate of 0.15, an accumulator value of 0.1, and a batch size of 16. For decoding, we use a beam of size 4.

The only component in LEAD that our proposed models change is the type of the RNN unit for the encoder/decoder. Namely, encRUM is a LEAD with a bidirectional RUM as an encoder (but with an LSTM decoder), decRUM is LEAD with a RUM as a decoder (but with a bi-LSTM encoder), and allRUM is LEAD with all LSTM units replaced by RUM ones. We train these models as LEAD, by minimizing the validation cross-entropy. We found that encRUM and allRUM take about 100k training steps to converge, while decRUM takes about 270k steps. Then, we turn on coverage training, as advised by See et al. (2017), and we train for a few thousand steps {2k,3k,4k,5k,8k}. The best ROUGE on dev was achieved for 2k steps, and this is what we used ultimately. We did not use time normalization as training was stable without it. We used the same hidden sizes for the LSTM, the RUM, and the mixed models. For the size of the hidden units, we tried {256, 360, 400, 512} on the dev set, and we found that 256 worked best overall.

Results.Table 4 shows ROUGE scores for the CNN / Daily Mail and the Science Daily test splits. We can see that RUM can easily replace LSTM in the pointer-generator network. We found that the best place to use RUM is in the decoder of the seq2seq model, since decRUM is better than encRUM and allRUM. Overall, we obtained the best results with decRUM 256 (lines 2 and 7), and we observed slight improvements for some ROUGE variants over previous work (i.e., with respect to lines 10–11).

We further trained decRUM with coverage for about 2,000 additional steps, which yielded 0.01 points of increase for ROUGE 1 (but with reduced ROUGE 2/L). We can conclude that here, as in the language modeling study (D), a combination of LSTM and RUM is better than using LSTM-only or RUM-only seq2seq models.

We conjecture that using RUM in the decoder is better because the encoder already has an attention mechanism and thus does not need much long-term memory, and would better focus on a more local context (as in LSTM). However, long-term memory is crucial for the decoder as it has to generate fluent output, and the attention mechanism cannot help it (i.e., better to use RUM). This is in line with our attention experiments on question answering. In future work, we plan to investigate combinations of LSTM and RUM units in more detail to identify optimal phase-coded attention.

Incorporating RUM into the seq2seq model yields larger gradients, compatible with stable training. Figure 6(a)shows the global norm of the gradients for our baseline models. Because of the $tanh$ activation, LSTM’s gradients hit the 1.0 baseline even though gradient clipping is 2.0. All RUM-based models have larger global norm. decRUM 360 sustains a slightly higher norm than LEAD, which might be beneficial. Panel 6(b), a consequence of 6(a), demonstrates that the RUM decoder sustains hidden states of higher norm throughout training. Panel 6(c) shows the contribution of the output at each encoder step to the gradient updates of the model. We observe that an LSTM encoder (in LEAD and decRUM) yields slightly higher gradient updates to the model, which is in line with our conjecture that it is better to use an LSTM encoder. Finally, panel 6(d) shows the gradient updates at each decoder step. Although the overall performance of LEAD and decRUM is similar, we note that the last few gradient updates from a RUM decoder are zero, while they are slightly above zero for LSTM. This happens because the target summaries for a minibatch are actually shorter than 100 tokens. Here, RUM exhibits an interesting property: It identifies that the target summary has ended, and thus for the subsequent extra steps, our model stops the gradients from updating the weights. An LSTM decoder keeps updating during the extra steps, which might indicate that it does not identify the end of the target summary properly.

We also compare our best decRUM 256 model to LEAD on the Science Daily data (lines 15–18). In Table 4, lines 15–17, we retrain the models from scratch. We can see that LEAD has clear advantage on the easiest task (line 15), which generally requires copying the first few sentences of the Science Daily article.

In line 16, this advantage decreases, as shuffling the paragraphs makes the task harder. We further observe that our RUM-based model demonstrates better performance on ROUGE F-2/L in line 17, where the task is highly abstractive.

In line 18, decRUM 256 and LEAD are pretrained on CNN / Daily Mail (models from lines 1–2), and our RUM-based model shows clear advantage on all ROUGE metrics. We also observe examples that are better than the ones coming from LEAD (see for example the story⁶ in Figure 1). We hypothesize that RUM is better on out-of-domain data due to its associative nature, as can be seen in Equation (2): At inference, the weight matrix updates for the hidden state depend explicitly on the current input.

We further test decRUM 256 and LEAD on the challenging task of producing popular summaries for research articles. The abundance of such articles online and the popular coverage of many of them (e.g., on Science Daily) provides an opportunity to develop models for automating science journalism.

The only directly related work⁷ is that of Vadapalli et al. (2018), who used research papers with corresponding popular style blog posts from Science Daily and phys.org, and aimed at generating the blog title. In their work, (i) they fed the paper title and its abstract into a heuristic function to extract relevant information, then (ii) they fed the output of this function into a pointer-generator network to produce a candidate title for the blog post.

Although we also use Science Daily and pointer- generator networks, we differ from the above work in a number of aspects. First, we focus on generating highlights, which are longer, more informative, and more complex than titles. Moreover, we feed the model a richer input, which includes not only the title and the abstract, but also the full text of the research paper.⁸ Finally, we skip (i), and in (ii) we encode for 1,000 steps (i.e., input words) and we decode for 100 steps. We observed that reading the first 1,000 words from the research paper is generally enough to generate a meaningful Science Daily-style highlight. Overall, we encode much more content from the research paper and we generate much longer highlights. To the best of our knowledge, our model is the only successful one in the domain of automatic science journalism that takes such a long input.

Figure 7 shows some highlights generated by our models, trained for 35k steps for decRUM and for 50k steps for LEAD. The highlights are grammatical, abstractive, and follow the Science Daily-style of reporting. The pointer-generator framework also allows for copying scientific terminology, which allows it to handle simultaneously domains ranging from computer science, to physics, to medicine. Interestingly, the words cancer and diseases are not mentioned in the research paper’s title or abstract, not even on the entire first page; yet, our models manage to extract them. See a demo and more examples in the link at footnote 1.

Here, we study the energy landscape of the loss function in order to give some intuition about why RUM’s choice of rotation is more appealing than what was used in previous phase-coded models. For simplicity, we only compare to GORU (Jing et al., 2017a) because GORU’s gated mechanism is most similar to that of RUM, and its orthogonal parametrization, given by Clements et al. (2016), is similar to that for the other orthogonal models in Section 2. Given a batch B = {b_i}_i, weights W = {w_j}_j, and a model F, the loss L(W,B) is defined as $∑jF(W,bj).$

In GORU, the weights are defined to be angles of rotations, and thus the summand is $F(W,bj)≡GORU(…,cos(wi),sin(wi),…,bj)$⁠. The arguments w_i of the trigonometric functions are independent of the batch element b_j, and all summands are in phase. Thus, the more trigonometric functions appear in F(W,b_j), the more local minima we expect to observe in L.

In contrast, for RUM we can write $F(W,bj)≡RUM(…,cos(g(wi,bj)),sin(g(wi,bj)),…,bj)$⁠, where g is the $arccos$ function that was used in defining the operation Rotation in Section 3. Because g depends on the input b_j, the summands F(W,b_j) are generally out of phase. As a result, L will not be close to periodic, which reduces the risk of falling into local minima.

We test our intuition by comparing the energy landscapes of RUM and GORU in Figure 8, following techniques similar to those in Li et al. (2018). For each model, we vary the weights in the orthogonal transformations: the Rotation operation for RUM, and the phase-coded kernel in GORU. Figure 8(a) and 8(c) show a 1D slice of the energy landscape. Note that 8(a) has less local minima than 8(c), which is also seen in Figure 8(b) and 8(d) for a 2D slice of the energy landscape.

We should be careful when using long-term memory RNN units if they are embedded in more complex networks (not just vanilla RNNs), such as stacked RNNs or seq2seq models with attention: Because such networks use unbounded activations (such as ReLU), the gradients could blow up in training. This is despite the unitary mechanism that stabilizes the vanilla RNN units. Along with the unitary models, RUM is also susceptible to blow-ups (as LSTM/GRU are), but it has a tunable mechanism solving this problem: time normalization.

We end this section with Table 5, which lists the best ingredients for successful RUM models.

We have proposed a representation unit for RNNs that combines properties of unitary learning and associative memory and enables really long-term memory modeling. We have further demonstrated that our model outperforms conventional RNNs on synthetic and on some real-world NLP tasks.

In future work, we plan to expand the representational power of our model by allowing λ in Equation (2) to be not only zero or one, but any real number.⁹ Second, we speculate that because our rotational matrix is a function of the RNN input (rather than being fixed after training, as in LSTM/GRU), RUM has a lot of potential for transfer learning. Finally, we would like to explore novel dataflows for RNN accelerators, which can run RUM efficiently.

We are very grateful to Dan Hogan from Science Daily for his help, to Daniel Dardani and Matthew Fucci for their advice, and to Thomas Frerix for the fruitful discussions.

This work was partially supported by the Army Research Office through the Institute for Soldier Nanotechnologies under contract W911NF-18- 2-0048; the National Science Foundation under grant no. CCF-1640012; and by the Semiconductor Research Corporation under grant no. 2016-EP-2693-B. This research is also supported in part by the MIT-SenseTime Alliance on Artificial Intelligence.