The past year has witnessed rapid advances in sequence-to-sequence (seq2seq) modeling for Machine Translation (MT). The classic RNN-based approaches to MT were first out-performed by the convolutional seq2seq model, which was then out-performed by the more recent Transformer model. Each of these new approaches consists of a fundamental architecture accompanied by a set of modeling and training techniques that are in principle applicable to other seq2seq architectures. In this paper, we tease apart the new architectures and their accompanying techniques in two ways. First, we identify several key modeling and training techniques, and apply them to the RNN architecture, yielding a new RNMT+ model that outperforms all of the three fundamental architectures on the benchmark WMT'14 English to French and English to German tasks. Second, we analyze the properties of each fundamental seq2seq architecture and devise new hybrid architectures intended to combine their strengths. Our hybrid models obtain further improvements, outperforming the RNMT+ model on both benchmark datasets.
The Transformer architecture has become increasingly popular over the past couple of years, owing to its impressive performance on a number of natural language processing (NLP) tasks. However, it may be argued that the Transformer architecture lacks an explicit hierarchical representation, as all computations occur on word-level representations alone, and therefore, learning structure poses a challenge for Transformer models. In the present work, we introduce hierarchical processing into the Transformer model, taking inspiration from the U-Net architecture, popular in computer vision for its hierarchical view of natural images. We propose a novel architecture that combines ideas from Transformer and U-Net models to incorporate hierarchy at multiple levels of abstraction. We empirically demonstrate that the proposed architecture outperforms the vanilla Transformer and strong baselines in the chit-chat dialogue and machine translation domains.
Self-attentive feed-forward sequence models have been shown to achieve impressive results on sequence modeling tasks, thereby presenting a compelling alternative to recurrent neural networks (RNNs) which has remained the de-facto standard architecture for many sequence modeling problems to date. Despite these successes, however, feed-forward sequence models like the Transformer fail to generalize in many tasks that recurrent models handle with ease (e.g. copying when the string lengths exceed those observed at training time). Moreover, and in contrast to RNNs, the Transformer model is not computationally universal, limiting its theoretical expressivity. In this paper we propose the Universal Transformer which addresses these practical and theoretical shortcomings and we show that it leads to improved performance on several tasks. Instead of recurring over the individual symbols of sequences like RNNs, the Universal Transformer repeatedly revises its representations of all symbols in the sequence with each recurrent step. In order to combine information from different parts of a sequence, it employs a self-attention mechanism in every recurrent step. Assuming sufficient memory, its recurrence makes the Universal Transformer computationally universal. We further employ an adaptive computation time (ACT) mechanism to allow the model to dynamically adjust the number of times the representation of each position in a sequence is revised. Beyond saving computation, we show that ACT can improve the accuracy of the model. Our experiments show that on various algorithmic tasks and a diverse set of large-scale language understanding tasks the Universal Transformer generalizes significantly better and outperforms both a vanilla Transformer and an LSTM in machine translation, and achieves a new state of the art on the bAbI linguistic reasoning task and the challenging LAMBADA language modeling task.
Music relies heavily on self-reference to build structure and meaning. We explore the Transformer architecture (Vaswani et al., 2017) as a generative model for music, as self-attention has shown compelling results on tasks that require long-term structure such as Wikipedia summary generation (Liu et al, 2018). However, timing information is critical for polyphonic music, and Transformer does not explicitly model absolute or relative timing in its structure. To address this challenge, Shaw et al. (2018) introduced relative position representations to self-attention to improve machine translation. However, the formulation was not scalable to longer sequences. We propose an improved formulation which reduces the memory requirements of the relative position computation from $O(l^2d)$ to $O(ld)$, making it possible to train much longer sequences and achieve faster convergence. In experiments on symbolic music we find that relative self-attention substantially improves sample quality for unconditioned generation and is able to generate sequences of lengths longer than those from the training set. When primed with an initial sequence, the model generates continuations that develop the prime coherently and exhibit long-term structure. Relative self-attention can be instrumental in capturing richer relationships within a musical piece.
Time series forecasting is an important problem across many domains, including predictions of solar plant energy output, electricity consumption, and traffic jam situation. In this paper, we propose to tackle such forecasting problem with Transformer. Although impressed by its performance in our preliminary study, we found its two major weaknesses: (1) locality-agnostics: the point-wise dot- product self-attention in canonical Transformer architecture is insensitive to local context, which can make the model prone to anomalies in time series; (2) memory bottleneck: space complexity of canonical Transformer grows quadratically with sequence length L, making directly modeling long time series infeasible. In order to solve these two issues, we first propose convolutional self-attention by producing queries and keys with causal convolution so that local context can be better incorporated into attention mechanism. Then, we propose LogSparse Transformer with only O(L(log L) 2) memory cost, improving forecasting accuracy for time series with fine granularity and strong long-term dependencies under constrained memory budget.