ICLR 2022 论文列表

Selecting an appropriate optimizer for a given problem is of major interest for researchers and practitioners. Many analytical optimizers have been proposed using a variety of theoretical and empirical approaches; however, none can offer a universal advantage over other competitive optimizers. We are thus motivated to study a new problem named Optimizer Amalgamation: how can we best combine a pool of "teacher" optimizers into a single "student" optimizer that can have stronger problem-specific performance? In this paper, we draw inspiration from the field of "learning to optimize" to use a learnable amalgamation target. First, we define three differentiable amalgamation mechanisms to amalgamate a pool of analytical optimizers by gradient descent. Then, in order to reduce variance of the amalgamation process, we also explore methods to stabilize the amalgamation process by perturbing the amalgamation target. Finally, we present experiments showing the superiority of our amalgamated optimizer compared to its amalgamated components and learning to optimize baselines, and the efficacy of our variance reducing perturbations.

There have been growing interests in leveraging experimental measurements to discover the underlying partial differential equations (PDEs) that govern complex physical phenomena. Although past research attempts have achieved great success in data-driven PDE discovery, the robustness of the existing methods cannot be guaranteed when dealing with low-quality measurement data. To overcome this challenge, we propose a novel physics-encoded discrete learning framework for discovering spatiotemporal PDEs from scarce and noisy data. The general idea is to (1) firstly introduce a novel deep convolutional-recurrent networks, which can encode prior physics knowledge (e.g., known terms, assumed PDE structure, initial/boundary conditions, etc.) while remaining flexible on representation capability, to accurately reconstruct high-fidelity data, and (2) then perform sparse regression with the reconstructed data to identify the analytical form of the governing PDEs. We validate our proposed framework on three high-dimensional PDE systems. The effectiveness and superiority of the proposed method over baselines are demonstrated.

The lottery ticket hypothesis states that a randomly-initialized neural network contains a small subnetwork which, when trained in isolation, can compete with the performance of the original network. Recent theoretical works proved an even stronger version: every sufficiently overparameterized (dense) neural network contains a subnetwork that, even without training, achieves accuracy comparable to that of the trained large network. These works left as an open problem to extend the result to convolutional neural networks (CNNs). In this work we provide such generalization by showing that, with high probability, it is possible to approximate any CNN by pruning a random CNN whose size is larger by a logarithmic factor.

Federated learning data is drawn from a distribution of distributions: clients are drawn from a meta-distribution, and their data are drawn from local data distributions. Generalization studies in federated learning should separate performance gaps from unseen client data (out-of-sample gap) from performance gaps from unseen client distributions (participation gap). In this work, we propose a framework for disentangling these performance gaps. Using this framework, we observe and explain differences in behavior across natural and synthetic federated datasets, indicating that dataset synthesis strategy can be important for realistic simulations of generalization in federated learning. We propose a semantic synthesis strategy that enables realistic simulation without naturally partitioned data. Informed by our findings, we call out community suggestions for future federated learning works.

Graph Neural Networks (GNNs) are gaining extensive attention for their application in graph data. However, the black-box nature of GNNs prevents users from understanding and trusting the models, thus hampering their applicability. Whereas explaining GNNs remains a challenge, most existing methods fall into approximation based and perturbation based approaches with suffer from faithfulness problems and unnatural artifacts respectively. To tackle these problems, we propose DEGREE (Decomposition based Explanation for GRaph nEural nEtworks) to provide a faithful explanation for GNN predictions. By decomposing the information generation and aggregation mechanism of GNNs, DEGREE allows tracking the contributions of specific components of the input graph to the final prediction. Based on this, we further design a subgraph level interpretation algorithm to reveal complex interactions between graph nodes that are overlooked by previous methods. The efficiency of our algorithm can be further improved by utilizing GNN characteristics. Finally, we conduct quantitative and qualitative experiments on synthetic and real-world datasets to demonstrate the effectiveness of DEGREE on node classification and graph classification tasks.

One of the challenges of graph-based semi-supervised learning over ordinary supervised learning for classification tasks lies in label utilization. The direct use of ground-truth labels in graphs for training purposes can result in a parametric model learning trivial degenerate solutions (e.g., an identity mapping from input to output). In addressing this issue, a label trick has recently been proposed in the literature and applied to a wide range of graph neural network (GNN) architectures, achieving state-of-the-art results on various datasets. The essential idea is to randomly split the observed labels on the graph and use a fraction of them as input to the model (along with original node features), and predict the remaining fraction. Despite its success in enabling GNNs to propagate features and labels simultaneously, this approach has never been analyzed from a theoretical perspective, nor fully explored across certain natural use cases. In this paper, we demonstrate that under suitable settings, this stochastic trick can be reduced to a more interpretable deterministic form, allowing us to better explain its behavior, including an emergent regularization effect, and motivate broader application scenarios. Our experimental results corroborate these analyses while also demonstrating improved node classification performance applying the label trick in new domains.

We perform approximate inference in state-space models with nonlinear state transitions. Without parameterizing a generative model, we apply Bayesian update formulas using a local linearity approximation parameterized by neural networks. It comes accompanied by a maximum likelihood objective that requires no supervision via uncorrupt observations or ground truth latent states. The optimization backpropagates through a recursion similar to the classical Kalman filter and smoother. Additionally, using an approximate conditional independence, we can perform smoothing without having to parameterize a separate model. In scientific applications, domain knowledge can give a linear approximation of the latent transition maps, which we can easily incorporate into our model. Usage of such domain knowledge is reflected in excellent results (despite our model's simplicity) on the chaotic Lorenz system compared to fully supervised and variational inference methods. Finally, we show competitive results on an audio denoising experiment.

Cross-domain object detection is more challenging than object classification since multiple objects exist in an image and the location of each object is unknown in the unlabeled target domain. As a result, when we adapt features of different objects to enhance the transferability of the detector, the features of the foreground and the background are easy to be confused, which may hurt the discriminability of the detector. Besides, previous methods focused on category adaptation but ignored another important part for object detection, i.e., the adaptation on bounding box regression. To this end, we propose D-adapt, namely Decoupled Adaptation, to decouple the adversarial adaptation and the training of the detector. Besides, we fill the blank of regression domain adaptation in object detection by introducing a bounding box adaptor. Experiments show that \textit{D-adapt} achieves state-of-the-art results on four cross-domain object detection tasks and yields 17\% and 21\% relative improvement on benchmark datasets Clipart1k and Comic2k in particular.

We study the dynamics of a neural network in function space when optimizing the mean squared error via gradient flow. We show that in the underparameterized regime the network learns eigenfunctions of an integral operator $T_K$ determined by the Neural Tangent Kernel at rates corresponding to their eigenvalues. For example, for uniformly distributed data on the sphere $S^{d - 1}$ and rotation invariant weight distributions, the eigenfunctions of $T_K$ are the spherical harmonics. Our results can be understood as describing a spectral bias in the underparameterized regime. The proofs use the concept of ``Damped Deviations'' where deviations of the NTK matter less for eigendirections with large eigenvalues. Aside from the underparameterized regime, the damped deviations point-of-view allows us to extend certain results in the literature in the overparameterized setting.

Transformer has become ubiquitous in sequence modeling tasks. As a key component of Transformer, self-attention does not scale to long sequences due to its quadratic time and space complexity with respect to the sequence length. To tackle this problem, recent work developed dynamic attention sparsification techniques based on Approximate Nearest Neighbor (ANN) methods, where similar queries and keys are allocated to the same hash bucket with high probability. However, the effectiveness of those ANN methods relies on the assumption that queries and keys should lie in the same space, which is not well justified. Besides, some of the ANN methods such as Locality-Sensitive Hashing (LSH) are randomized and cannot fully utilize the available real data distributions. To overcome these issues, this paper proposes a new strategy for sparse attention, namely LHA (Learning-to-Hash Attention), which directly learns separate parameterized hash functions for queries and keys, respectively. Another advantage of LHA is that it does not impose extra constraints for queries and keys, which makes it applicable to the wide range of pre-trained Transformer models. Our experiments on evaluation of the WikiText-103 dataset for language modeling, the GLUE benchmark for natural language understanding, and the Lang-Range-Arena benchmark for multiple tasks (text/image classification, retrieval, etc.) show the superior performance of LHA over other strong Transformer variants.

Diffusion models have emerged as an expressive family of generative models rivaling GANs in sample quality and autoregressive models in likelihood scores. Standard diffusion models typically require hundreds of forward passes through the model to generate a single high-fidelity sample. We introduce Differentiable Diffusion Sampler Search (DDSS): a method that optimizes fast samplers for any pre-trained diffusion model by differentiating through sample quality scores. We also present Generalized Gaussian Diffusion Models (GGDM), a family of flexible non-Markovian samplers for diffusion models. We show that optimizing the degrees of freedom of GGDM samplers by maximizing sample quality scores via gradient descent leads to improved sample quality. Our optimization procedure backpropagates through the sampling process using the reparametrization trick and gradient rematerialization. DDSS achieves strong results on unconditional image generation across various datasets (e.g., FID scores on LSUN church 128x128 of 11.6 with only 10 inference steps, and 4.82 with 20 steps, compared to 51.1 and 14.9 with strongest DDPM/DDIM baselines). Our method is compatible with any pre-trained diffusion model without fine-tuning or re-training required.

Random pruning is arguably the most naive way to attain sparsity in neural networks, but has been deemed uncompetitive by either post-training pruning or sparse training. In this paper, we focus on sparse training and highlight a perhaps counter-intuitive finding, that random pruning at initialization can be quite powerful for the sparse training of modern neural networks. Without any delicate pruning criteria or carefully pursued sparsity structures, we empirically demonstrate that sparsely training a randomly pruned network from scratch can match the performance of its dense equivalent. There are two key factors that contribute to this revival: (i) $the network sizes matter$: as the original dense networks grow wider and deeper, the performance of training a randomly pruned sparse network will quickly grow to matching that of its dense equivalent, even at high sparsity ratios; (ii) $appropriate layer-wise sparsity ratios$ can be pre-chosen for sparse training, which shows to be another important performance booster. Simple as it looks, a randomly pruned subnetwork of Wide ResNet-50 can be sparsely trained to outperforming a dense Wide ResNet-50, on ImageNet. We also observed such randomly pruned networks outperform dense counterparts in other favorable aspects, such as out-of-distribution detection, uncertainty estimation, and adversarial robustness. Overall, our results strongly suggest there is larger-than-expected room for sparse training at scale, and the benefits of sparsity might be more universal beyond carefully designed pruning. Our source code can be found at https://github.com/VITA-Group/Random_Pruning.

Recent developments in large-scale machine learning suggest that by scaling up data, model size and training time properly, one might observe that improvements in pre-training would transfer favorably to most downstream tasks. In this work we systematically study this phenomena and establish that, as we increase the upstream accuracy, performance of downstream tasks \emph{saturates}. In particular, we investigate more than 4800 experiments on Vision Transformers, MLP-Mixers and ResNets with number of parameters ranging from ten million to ten billion, trained on the largest scale of available image data (JFT, ImageNet21K) and evaluated on more than 20 downstream image recognition tasks. We propose a model for downstream performance that reflects the saturation phenomena and captures the nonlinear relationship in performance of upstream and downstream tasks. Delving deeper to understand the reasons that give rise to these phenomena, we show that the observed saturation behavior is closely related to the way that representations evolve through the layers of the models. We showcase an even more extreme scenario where performance on upstream and downstream are at odds with each other. That is, in order to have a better downstream performance, we need to hurt upstream accuracy.

Transformers are neural network architectures that achieve remarkable performance in many areas. However, the core component of Transformers, multi-head self-attention (MHSA), is mainly derived from heuristics, and the interactions across its components are not well understood. To address the problem, we first introduce a mathematically rigorous and yet intuitive tensor diagram representation of MHSA. Guided by tensor diagram representations, we propose a novel design, namely Tunable Transformers (Tuformers), by allowing data-driven weights across heads, whereas MHSA adopts pre-defined and fixed weights across heads, as will be explained in our paper. Tuformers naturally reveal a flexible design space that a user, depending on the needs, can choose a structure that has either improved performance (generalization error) or higher model efficiency. Any pre-trained Transformer can be an initialization of the corresponding Tuformer with trainable number of heads for efficient training and fine-tuning. Tuformers universally outperform Transformers on various tasks across multiple domains under a wide range of model sizes.

How can we measure the reasoning capabilities of intelligence systems? Visual question answering provides a convenient framework for testing the model's abilities by interrogating the model through questions about the scene. However, despite scores of various visual QA datasets and architectures, which sometimes yield even a super-human performance, the question of whether those architectures can actually reason remains open to debate. To answer this, we extend the visual question answering framework and propose the following behavioral test in the form of a two-player game. We consider black-box neural models of CLEVR. These models are trained on a diagnostic dataset benchmarking reasoning. Next, we train an adversarial player that re-configures the scene to fool the CLEVR model. We show that CLEVR models, which otherwise could perform at a ``human-level'', can easily be fooled by our agent. Our results put in doubt whether data-driven approaches can do reasoning without exploiting the numerous biases that are often present in those datasets. Finally, we also propose a controlled experiment measuring the efficiency of such models to learn and perform reasoning.

Musical expression requires control of both what notes that are played, and how they are performed. Conventional audio synthesizers provide detailed expressive controls, but at the cost of realism. Black-box neural audio synthesis and concatenative samplers can produce realistic audio, but have few mechanisms for control. In this work, we introduce MIDI-DDSP a hierarchical model of musical instruments that enables both realistic neural audio synthesis and detailed user control. Starting from interpretable Differentiable Digital Signal Processing (DDSP) synthesis parameters, we infer musical notes and high-level properties of their expressive performance (such as timbre, vibrato, dynamics, and articulation). This creates a 3-level hierarchy (notes, performance, synthesis) that affords individuals the option to intervene at each level, or utilize trained priors (performance given notes, synthesis given performance) for creative assistance. Through quantitative experiments and listening tests, we demonstrate that this hierarchy can reconstruct high-fidelity audio, accurately predict performance attributes for a note sequence, independently manipulate the attributes of a given performance, and as a complete system, generate realistic audio from a novel note sequence. By utilizing an interpretable hierarchy, with multiple levels of granularity, MIDI-DDSP opens the door to assistive tools to empower individuals across a diverse range of musical experience.

Recent research has shown that numerous human-interpretable directions exist in the latent space of GANs. In this paper, we develop an automatic procedure for finding directions that lead to foreground-background image separation, and we use these directions to train an image segmentation model without human supervision. Our method is generator-agnostic, producing strong segmentation results with a wide range of different GAN architectures. Furthermore, by leveraging GANs pretrained on large datasets such as ImageNet, we are able to segment images from a range of domains without further training or finetuning. Evaluating our method on image segmentation benchmarks, we compare favorably to prior work while using neither human supervision nor access to the training data. Broadly, our results demonstrate that automatically extracting foreground-background structure from pretrained deep generative models can serve as a remarkably effective substitute for human supervision.

This paper considers minimax optimization $\min_x \max_y f(x, y)$ in the challenging setting where $f$ can be both nonconvex in $x$ and nonconcave in $y$. Though such optimization problems arise in many machine learning paradigms including training generative adversarial networks (GANs) and adversarially robust models, from a theoretical point of view, two fundamental issues remain: (i) the absence of simple and efficiently computable optimality notions, and (ii) cyclic or diverging behavior of existing algorithms. This paper proposes a new theoretical framework for nonconvex-nonconcave minimax optimization that addresses both of the above issues. The starting point of this paper is the observation that, under a computational budget, the max-player can not fully maximize $f(x,\cdot)$ since nonconcave maximization is NP-hard in general. So, we propose a new framework, and a corresponding algorithm, for the min-player to play against \emph{smooth algorithms} deployed by the adversary (i.e., the max-player) instead of against full maximization. Our algorithm is guaranteed to make monotonic progress (thus having no limit cycles or diverging behavior), and to find an appropriate ``stationary point'' in a polynomial number of iterations. Our framework covers practically relevant settings where the smooth algorithms deployed by the adversary are multi-step stochastic gradient ascent, and its accelerated version. We further present experimental results that confirm our theoretical findings and demonstrate the effectiveness of the proposed approach in practice on simple, conceptual settings.

The practice of data augmentation has been extensively used to boost the performance of deep neural networks for various NLP tasks. It is more effective when only a limited number of labeled samples is available, e.g., low-data or class-imbalanced regimes. Most current augmentation techniques rely on parameter tuning or inherent randomness; hence, their effectiveness largely varies on the tasks. To efficiently find the best augmentation strategy for each task, learning data augmentation policy is a promising solution, but the question of what makes a good augmentation in NLP tasks and how to design the reward function for learning a good policy remains under-explored. To answer this, we hypothesize that good data augmentation should construct more diverse and challenging samples for providing informative training signals, while avoiding the risk of losing the semantics of original samples. Therefore, we design a novel reward function for updating the augmentation policy to construct difficult but not too different samples (DND). Particularly, we jointly optimize a data augmentation policy while training the model, to construct the augmented samples with low confidence but a high semantic similarity with original ones. In addition, we introduce a sample re-weighting scheme to focus on difficult augmented samples after the original ones are learned confidently for more effective learning from the augmented ones. Our learning-based augmentation outperforms the recent state-of-the-art augmentation schemes on various text classification tasks and GLUE benchmark by successfully discovering the effective augmentations for each task. Remarkably, our method is more effective on the challenging low-data and class-imbalanced regimes, and the learned augmentation policy is well-transferable to the different tasks and models.