ICLR 2020 论文列表

Learning to disentangle the hidden factors of variations within a set of observations is a key task for artificial intelligence. We present a unified formulation for class and content disentanglement and use it to illustrate the limitations of current methods. We therefore introduce LORD, a novel method based on Latent Optimization for Representation Disentanglement. We find that latent optimization, along with an asymmetric noise regularization, is superior to amortized inference for achieving disentangled representations. In extensive experiments, our method is shown to achieve better disentanglement performance than both adversarial and non-adversarial methods that use the same level of supervision. We further introduce a clustering-based approach for extending our method for settings that exhibit in-class variation with promising results on the task of domain translation.

The recently proposed BERT (Devlin et al., 2019) has shown great power on a variety of natural language understanding tasks, such as text classification, reading comprehension, etc. However, how to effectively apply BERT to neural machine translation (NMT) lacks enough exploration. While BERT is more commonly used as fine-tuning instead of contextual embedding for downstream language understanding tasks, in NMT, our preliminary exploration of using BERT as contextual embedding is better than using for fine-tuning. This motivates us to think how to better leverage BERT for NMT along this direction. We propose a new algorithm named BERT-fused model, in which we first use BERT to extract representations for an input sequence, and then the representations are fused with each layer of the encoder and decoder of the NMT model through attention mechanisms. We conduct experiments on supervised (including sentence-level and document-level translations), semi-supervised and unsupervised machine translation, and achieve state-of-the-art results on seven benchmark datasets. Our code is available at https://github.com/bert-nmt/bert-nmt

We propose order learning to determine the order graph of classes, representing ranks or priorities, and classify an object instance into one of the classes. To this end, we design a pairwise comparator to categorize the relationship between two instances into one of three cases: one instance is `greater than,' `similar to,' or `smaller than' the other. Then, by comparing an input instance with reference instances and maximizing the consistency among the comparison results, the class of the input can be estimated reliably. We apply order learning to develop a facial age estimator, which provides the state-of-the-art performance. Moreover, the performance is further improved when the order graph is divided into disjoint chains using gender and ethnic group information or even in an unsupervised manner.

Neural Architecture Search (NAS) is an exciting new field which promises to be as much as a game-changer as Convolutional Neural Networks were in 2012. Despite many great works leading to substantial improvements on a variety of tasks, comparison between different methods is still very much an open issue. While most algorithms are tested on the same datasets, there is no shared experimental protocol followed by all. As such, and due to the under-use of ablation studies, there is a lack of clarity regarding why certain methods are more effective than others. Our first contribution is a benchmark of 8 NAS methods on 5 datasets. To overcome the hurdle of comparing methods with different search spaces, we propose using a method’s relative improvement over the randomly sampled average architecture, which effectively removes advantages arising from expertly engineered search spaces or training protocols. Surprisingly, we find that many NAS techniques struggle to significantly beat the average architecture baseline. We perform further experiments with the commonly used DARTS search space in order to understand the contribution of each component in the NAS pipeline. These experiments highlight that: (i) the use of tricks in the evaluation protocol has a predominant impact on the reported performance of architectures; (ii) the cell-based search space has a very narrow accuracy range, such that the seed has a considerable impact on architecture rankings; (iii) the hand-designed macrostructure (cells) is more important than the searched micro-structure (operations); and (iv) the depth-gap is a real phenomenon, evidenced by the change in rankings between 8 and 20 cell architectures. To conclude, we suggest best practices, that we hope will prove useful for the community and help mitigate current NAS pitfalls, e.g. difficulties in reproducibility and comparison of search methods. The code used is available at https://github.com/antoyang/NAS-Benchmark.

In this paper, we consider the problem of training neural networks (NN). To promote a NN with specific structures, we explicitly take into consideration the nonsmooth regularization (such as L1-norm) and constraints (such as interval constraint). This is formulated as a constrained nonsmooth nonconvex optimization problem, and we propose a convergent proximal-type stochastic gradient descent (Prox-SGD) algorithm. We show that under properly selected learning rates, momentum eventually resembles the unknown real gradient and thus is crucial in analyzing the convergence. We establish that with probability 1, every limit point of the sequence generated by the proposed Prox-SGD is a stationary point. Then the Prox-SGD is tailored to train a sparse neural network and a binary neural network, and the theoretical analysis is also supported by extensive numerical tests.

Natural question generation (QG) aims to generate questions from a passage and an answer. Previous works on QG either (i) ignore the rich structure information hidden in text, (ii) solely rely on cross-entropy loss that leads to issues like exposure bias and inconsistency between train/test measurement, or (iii) fail to fully exploit the answer information. To address these limitations, in this paper, we propose a reinforcement learning (RL) based graph-to-sequence (Graph2Seq) model for QG. Our model consists of a Graph2Seq generator with a novel Bidirectional Gated Graph Neural Network based encoder to embed the passage, and a hybrid evaluator with a mixed objective combining both cross-entropy and RL losses to ensure the generation of syntactically and semantically valid text. We also introduce an effective Deep Alignment Network for incorporating the answer information into the passage at both the word and contextual levels. Our model is end-to-end trainable and achieves new state-of-the-art scores, outperforming existing methods by a significant margin on the standard SQuAD benchmark.

Recent theoretical work has guaranteed that overparameterized networks trained by gradient descent achieve arbitrarily low training error, and sometimes even low test error. The required width, however, is always polynomial in at least one of the sample size $n$, the (inverse) target error $1/\epsilon$, and the (inverse) failure probability $1/\delta$. This work shows that $\widetilde{\Theta}(1/\epsilon)$ iterations of gradient descent with $\widetilde{\Omega}(1/\epsilon^2)$ training examples on two-layer ReLU networks of any width exceeding $\textrm{polylog}(n,1/\epsilon,1/\delta)$ suffice to achieve a test misclassification error of $\epsilon$. We also prove that stochastic gradient descent can achieve $\epsilon$ test error with polylogarithmic width and $\widetilde{\Theta}(1/\epsilon)$ samples. The analysis relies upon the separation margin of the limiting kernel, which is guaranteed positive, can distinguish between true labels and random labels, and can give a tight sample-complexity analysis in the infinite-width setting.

Quantum computing is a powerful computational paradigm with applications in several fields, including machine learning. In the last decade, deep learning, and in particular Convolutional Neural Networks (CNN), have become essential for applications in signal processing and image recognition. Quantum deep learning, however, remains a challenging problem, as it is difficult to implement non linearities with quantum unitaries. In this paper we propose a quantum algorithm for evaluating and training deep convolutional neural networks with potential speedups over classical CNNs for both the forward and backward passes. The quantum CNN (QCNN) reproduces completely the outputs of the classical CNN and allows for non linearities and pooling operations. The QCNN is in particular interesting for deep networks and could allow new frontiers in the image recognition domain, by allowing for many more convolution kernels, larger kernels, high dimensional inputs and high depth input channels. We also present numerical simulations for the classification of the MNIST dataset to provide practical evidence for the efficiency of the QCNN.

Temporal point processes are the dominant paradigm for modeling sequences of events happening at irregular intervals. The standard way of learning in such models is by estimating the conditional intensity function. However, parameterizing the intensity function usually incurs several trade-offs. We show how to overcome the limitations of intensity-based approaches by directly modeling the conditional distribution of inter-event times. We draw on the literature on normalizing flows to design models that are flexible and efficient. We additionally propose a simple mixture model that matches the flexibility of flow-based models, but also permits sampling and computing moments in closed form. The proposed models achieve state-of-the-art performance in standard prediction tasks and are suitable for novel applications, such as learning sequence embeddings and imputing missing data.

Experimental reproducibility and replicability are critical topics in machine learning. Authors have often raised concerns about their lack in scientific publications to improve the quality of the field. Recently, the graph representation learning field has attracted the attention of a wide research community, which resulted in a large stream of works. As such, several Graph Neural Network models have been developed to effectively tackle graph classification. However, experimental procedures often lack rigorousness and are hardly reproducible. Motivated by this, we provide an overview of common practices that should be avoided to fairly compare with the state of the art. To counter this troubling trend, we ran more than 47000 experiments in a controlled and uniform framework to re-evaluate five popular models across nine common benchmarks. Moreover, by comparing GNNs with structure-agnostic baselines we provide convincing evidence that, on some datasets, structural information has not been exploited yet. We believe that this work can contribute to the development of the graph learning field, by providing a much needed grounding for rigorous evaluations of graph classification models.

We propose a learned image-guided rendering technique that combines the benefits of image-based rendering and GAN-based image synthesis. The goal of our method is to generate photo-realistic re-renderings of reconstructed objects for virtual and augmented reality applications (e.g., virtual showrooms, virtual tours and sightseeing, the digital inspection of historical artifacts). A core component of our work is the handling of view-dependent effects. Specifically, we directly train an object-specific deep neural network to synthesize the view-dependent appearance of an object. As input data we are using an RGB video of the object. This video is used to reconstruct a proxy geometry of the object via multi-view stereo. Based on this 3D proxy, the appearance of a captured view can be warped into a new target view as in classical image-based rendering. This warping assumes diffuse surfaces, in case of view-dependent effects, such as specular highlights, it leads to artifacts. To this end, we propose EffectsNet, a deep neural network that predicts view-dependent effects. Based on these estimations, we are able to convert observed images to diffuse images. These diffuse images can be projected into other views. In the target view, our pipeline reinserts the new view-dependent effects. To composite multiple reprojected images to a final output, we learn a composition network that outputs photo-realistic results. Using this image-guided approach, the network does not have to allocate capacity on ``remembering'' object appearance, instead it learns how to combine the appearance of captured images. We demonstrate the effectiveness of our approach both qualitatively and quantitatively on synthetic as well as on real data.

Simultaneous machine translation models start generating a target sequence before they have encoded or read the source sequence. Recent approach for this task either apply a fixed policy on transformer, or a learnable monotonic attention on a weaker recurrent neural network based structure. In this paper, we propose a new attention mechanism, Monotonic Multihead Attention (MMA), which introduced the monotonic attention mechanism to multihead attention. We also introduced two novel interpretable approaches for latency control that are specifically designed for multiple attentions. We apply MMA to the simultaneous machine translation task and demonstrate better latency-quality tradeoffs compared to MILk, the previous state-of-the-art approach.

When performing imitation learning from expert demonstrations, distribution matching is a popular approach, in which one alternates between estimating distribution ratios and then using these ratios as rewards in a standard reinforcement learning (RL) algorithm. Traditionally, estimation of the distribution ratio requires on-policy data, which has caused previous work to either be exorbitantly data- inefficient or alter the original objective in a manner that can drastically change its optimum. In this work, we show how the original distribution ratio estimation objective may be transformed in a principled manner to yield a completely off-policy objective. In addition to the data-efficiency that this provides, we are able to show that this objective also renders the use of a separate RL optimization unnecessary. Rather, an imitation policy may be learned directly from this objective without the use of explicit rewards. We call the resulting algorithm ValueDICE and evaluate it on a suite of popular imitation learning benchmarks, finding that it can achieve state-of-the-art sample efficiency and performance.

As deep neural networks (DNNs) achieve tremendous success across many application domains, researchers tried to explore in many aspects on why they generalize well. In this paper, we provide a novel perspective on these issues using the gradient signal to noise ratio (GSNR) of parameters during training process of DNNs. The GSNR of a parameter is simply defined as the ratio between its gradient's squared mean and variance, over the data distribution. Based on several approximations, we establish a quantitative relationship between model parameters' GSNR and the generalization gap. This relationship indicates that larger GSNR during training process leads to better generalization performance. Futher, we show that, different from that of shallow models (e.g. logistic regression, support vector machines), the gradient descent optimization dynamics of DNNs naturally produces large GSNR during training, which is probably the key to DNNs’ remarkable generalization ability.

The ability to represent and compare machine learning models is crucial in order to quantify subtle model changes, evaluate generative models, and gather insights on neural network architectures. Existing techniques for comparing data distributions focus on global data properties such as mean and covariance; in that sense, they are extrinsic and uni-scale. We develop a first-of-its-kind intrinsic and multi-scale method for characterizing and comparing data manifolds, using a lower-bound of the spectral variant of the Gromov-Wasserstein inter-manifold distance, which compares all data moments. In a thorough experimental study, we demonstrate that our method effectively discerns the structure of data manifolds even on unaligned data of different dimensionalities; moreover, we showcase its efficacy in evaluating the quality of generative models.

Adversarial patch attacks are among one of the most practical threat models against real-world computer vision systems. This paper studies certified and empirical defenses against patch attacks. We begin with a set of experiments showing that most existing defenses, which work by pre-processing input images to mitigate adversarial patches, are easily broken by simple white-box adversaries. Motivated by this finding, we propose the first certified defense against patch attacks, and propose faster methods for its training. Furthermore, we experiment with different patch shapes for testing, obtaining surprisingly good robustness transfer across shapes, and present preliminary results on certified defense against sparse attacks. Our complete implementation can be found on: https://github.com/Ping-C/certifiedpatchdefense.

In this work we study locality and compositionality in the context of learning representations for Zero Shot Learning (ZSL). In order to well-isolate the importance of these properties in learned representations, we impose the additional constraint that, differently from most recent work in ZSL, no pre-training on different datasets (e.g. ImageNet) is performed. The results of our experiment show how locality, in terms of small parts of the input, and compositionality, i.e. how well can the learned representations be expressed as a function of a smaller vocabulary, are both deeply related to generalization and motivate the focus on more local-aware models in future research directions for representation learning.

In this paper, we introduce Mamus for constructing multilingual sparse word representations. Our algorithm operates by determining a shared set of semantic units which get reutilized across languages, providing it a competitive edge both in terms of speed and evaluation performance. We demonstrate that our proposed algorithm behaves competitively to strong baselines through a series of rigorous experiments performed towards downstream applications spanning over dependency parsing, document classification and natural language inference. Additionally, our experiments relying on the QVEC-CCA evaluation score suggests that the proposed sparse word representations convey an increased interpretability as opposed to alternative approaches. Finally, we are releasing our multilingual sparse word representations for the 27 typologically diverse set of languages that we conducted our various experiments on.

Predicting outcomes and planning interactions with the physical world are long-standing goals for machine learning. A variety of such tasks involves continuous physical systems, which can be described by partial differential equations (PDEs) with many degrees of freedom. Existing methods that aim to control the dynamics of such systems are typically limited to relatively short time frames or a small number of interaction parameters. We present a novel hierarchical predictor-corrector scheme which enables neural networks to learn to understand and control complex nonlinear physical systems over long time frames. We propose to split the problem into two distinct tasks: planning and control. To this end, we introduce a predictor network that plans optimal trajectories and a control network that infers the corresponding control parameters. Both stages are trained end-to-end using a differentiable PDE solver. We demonstrate that our method successfully develops an understanding of complex physical systems and learns to control them for tasks involving PDEs such as the incompressible Navier-Stokes equations.