ICLR 2022 论文列表

Goal-oriented Reinforcement Learning (GoRL) is a promising approach for scaling up RL techniques on sparse reward environments requiring long horizon planning. Recent works attempt to build suitable abstraction graph of the environment and enhance GoRL with classical graphical methods such as shortest path searching; however, these approaches mainly focus on either graph construction or agent exploitation, but leave the exploration lack of study. This paper proposes Graph-enhanced GoRL (G2RL), a new GoRL framework for effective exploration and efficient training based on the state-transition graph. We first introduce the optimal goals for exploration on the graph and then use them as supervised signals to train the goal generator in G2RL in a hindsight manner. Furthermore, we define relevant trajectories of a state based on its graph neighborhood and show that giving high priority to these trajectories would lead to an efficient policy learning. In addition to the theoretical results regarding optimal goal generation, our empirical results on standard discrete and continuous control benchmarks show that leveraging the state-transition graph is beneficial for GoRL to learn an effective and informative exploration strategy and outperform the state-of-the-art methods.

We introduce an adaptive tree search algorithm, which is a deterministic variant of Monte Carlo tree search, that can find high-scoring outputs under translation models that make no assumptions about the form or structure of the search objective. This algorithm enables the exploration of new kinds of models that are unencumbered by constraints imposed to make decoding tractable, such as autoregressivity or conditional independence assumptions. When applied to autoregressive models, our algorithm has different biases than beam search has, which enables a new analysis of the role of decoding bias in autoregressive models. Empirically, we show that our adaptive tree search algorithm finds outputs with substantially better model scores compared to beam search in autoregressive models, and compared to reranking techniques in models whose scores do not decompose additively with respect to the words in the output. We also characterise the correlation of several translation model objectives with respect to BLEU. We find that while some standard models are poorly calibrated and benefit from the beam search bias, other often more robust models (autoregressive models tuned to maximize expected automatic metric scores, the noisy channel model and a newly proposed objective) benefit from increasing amounts of search using our proposed decoder, whereas the beam search bias limits the improvements obtained from such objectives. Thus, we argue that as models improve, the improvements may be masked by over-reliance on beam search or reranking based methods.

Data augmentation is an effective way to improve the generalization capability of modern deep learning models. However, the underlying augmentation methods mostly rely on handcrafted operations. Moreover, an augmentation policy useful to one dataset may not transfer well to other datasets. Therefore, Automated Data Augmentation (AutoDA) methods, like \textit{AutoAugment} and \textit{Population-based Augmentation}, have been proposed recently to automate the process of searching for optimal augmentation policies. However, the augmentation policies found are not adaptive to the dataset used, hindering the effectiveness of these AutoDA methods. In this paper, we propose a novel AutoDA method called \texttt{AdaAug} to efficiently learn adaptive augmentation policies in a class-dependent and potentially instance-dependent manner. Our experiments show that the adaptive augmentation policies learned by our method transfer well to unseen datasets such as the Oxford Flowers, Oxford-IIT Pets, FGVC Aircraft, and Stanford Cars datasets when compared with other AutoDA baselines. In addition, our method also achieves state-of-the-art performance on the CIFAR-10, CIFAR-100, and SVHN datasets.

We address the problem of building agents whose goal is to learn to execute out-of distribution (OOD) multi-task instructions expressed in temporal logic (TL) by using deep reinforcement learning (DRL). Recent works provided evidence that the agent's neural architecture is a key feature when DRL agents are learning to solve OOD tasks in TL. Yet, the studies on this topic are still in their infancy. In this work, we propose a new deep learning configuration with inductive biases that lead agents to generate latent representations of their current goal, yielding a stronger generalization performance. We use these latent-goal networks within a neuro-symbolic framework that executes multi-task formally-defined instructions and contrast the performance of the proposed neural networks against employing different state-of-the-art (SOTA) architectures when generalizing to unseen instructions in OOD environments.

Learning rate schedulers have been widely adopted in training deep neural networks. Despite their practical importance, there is a discrepancy between its practice and its theoretical analysis. For instance, it is not known what schedules of SGD achieve best convergence, even for simple problems such as optimizing quadratic objectives. In this paper, we propose Eigencurve, the first family of learning rate schedules that can achieve minimax optimal convergence rates (up to a constant) for SGD on quadratic objectives when the eigenvalue distribution of the underlying Hessian matrix is skewed. The condition is quite common in practice. Experimental results show that Eigencurve can significantly outperform step decay in image classification tasks on CIFAR-10, especially when the number of epochs is small. Moreover, the theory inspires two simple learning rate schedulers for practical applications that can approximate eigencurve. For some problems, the optimal shape of the proposed schedulers resembles that of cosine decay, which sheds light to the success of cosine decay for such situations. For other situations, the proposed schedulers are superior to cosine decay.

Generalization is one of the fundamental issues in machine learning. However, traditional techniques like uniform convergence may be unable to explain generalization under overparameterization \citep{nagarajan2019uniform}. As alternative approaches, techniques based on stability analyze the training dynamics and derive algorithm-dependent generalization bounds. Unfortunately, the stability-based bounds are still far from explaining the surprising generalization in deep learning since neural networks usually suffer from unsatisfactory stability. This paper proposes a novel decomposition framework to improve the stability-based bounds via a more fine-grained analysis of the signal and noise, inspired by the observation that neural networks converge relatively slowly when fitting noise (which indicates better stability). Concretely, we decompose the excess risk dynamics and apply the stability-based bound only on the noise component. The decomposition framework performs well in both linear regimes (overparameterized linear regression) and non-linear regimes (diagonal matrix recovery). Experiments on neural networks verify the utility of the decomposition framework.

Using a single value function (critic) shared over multiple tasks in Actor-Critic multi-task reinforcement learning (MTRL) can result in negative interference between tasks, which can compromise learning performance. Multi-Critic Actor Learning (MultiCriticAL) proposes instead maintaining separate critics for each task being trained while training a single multi-task actor. Explicitly distinguishing between tasks also eliminates the need for critics to learn to do so and mitigates interference between task-value estimates. MultiCriticAL is tested in the context of multi-style learning, a special case of MTRL where agents are trained to behave with different distinct behavior styles, and yields up to 56% performance gains over the single-critic baselines and even successfully learns behavior styles in cases where single-critic approaches may simply fail to learn. In a simulated real-world use case, MultiCriticAL enables learning policies that smoothly transition between multiple fighting styles on an experimental build of EA’s UFC game.

Dictionary learning consists of finding a sparse representation from noisy data and is a common way to encode data-driven prior knowledge on signals. Alternating minimization (AM) is standard for the underlying optimization, where gradient descent steps alternate with sparse coding procedures. The major drawback of this method is its prohibitive computational cost, making it unpractical on large real-world data sets. This work studies an approximate formulation of dictionary learning based on unrolling and compares it to alternating minimization to find the best trade-off between speed and precision. We analyze the asymptotic behavior and convergence rate of gradients estimates in both methods. We show that unrolling performs better on the support of the inner problem solution and during the first iterations. Finally, we apply unrolling on pattern learning in magnetoencephalography (MEG) with the help of a stochastic algorithm and compare the performance to a state-of-the-art method.

What is the state of the art in continual machine learning? Although a natural question for predominant static benchmarks, the notion to train systems in a lifelong manner entails a plethora of additional challenges with respect to set-up and evaluation. The latter have recently sparked a growing amount of critiques on prominent algorithm-centric perspectives and evaluation protocols being too narrow, resulting in several attempts at constructing guidelines in favor of specific desiderata or arguing against the validity of prevalent assumptions. In this work, we depart from this mindset and argue that the goal of a precise formulation of desiderata is an ill-posed one, as diverse applications may always warrant distinct scenarios. Instead, we introduce the Continual Learning EValuation Assessment Compass: the CLEVA-Compass. The compass provides the visual means to both identify how approaches are practically reported and how works can simultaneously be contextualized in the broader literature landscape. In addition to promoting compact specification in the spirit of recent replication trends, it thus provides an intuitive chart to understand the priorities of individual systems, where they resemble each other, and what elements are missing towards a fair comparison.

Graph self-supervised learning has gained increasing attention due to its capacity to learn expressive node representations. Many pretext tasks, or loss functions have been designed from distinct perspectives. However, we observe that different pretext tasks affect downstream tasks differently cross datasets, which suggests that searching pretext tasks is crucial for graph self-supervised learning. Different from existing works focusing on designing single pretext tasks, this work aims to investigate how to automatically leverage multiple pretext tasks effectively. Nevertheless, evaluating representations derived from multiple pretext tasks without direct access to ground truth labels makes this problem challenging. To address this obstacle, we make use of a key principle of many real-world graphs, i.e., homophily, or the principle that ``like attracts like,'' as the guidance to effectively search various self-supervised pretext tasks. We provide theoretical understanding and empirical evidence to justify the flexibility of homophily in this search task. Then we propose the AutoSSL framework which can automatically search over combinations of various self-supervised tasks. By evaluating the framework on 7 real-world datasets, our experimental results show that AutoSSL can significantly boost the performance on downstream tasks including node clustering and node classification compared with training under individual tasks.

Searching for the architecture cells is a dominant paradigm in NAS. However, little attention has been devoted to the analysis of the cell-based search spaces even though it is highly important for the continual development of NAS. In this work, we conduct an empirical post-hoc analysis of architectures from the popular cell-based search spaces and find that the existing search spaces contain a high degree of redundancy: the architecture performance is less sensitive to changes at large parts of the cells, and universally adopted design rules, like the explicit search for a reduction cell, significantly increase the complexities but have very limited impact on the performance. Across architectures found by a diverse set of search strategies, we consistently find that the parts of the cells that do matter for architecture performance often follow similar and simple patterns. By constraining cells to include these patterns, randomly sampled architectures can match or even outperform the state of the art. These findings cast doubts into our ability to discover truly novel architectures in the existing cell-based search spaces and, inspire our suggestions for improvement to guide future NAS research. Code is available at https://github.com/xingchenwan/cell-based-NAS-analysis.

Deep Generative Networks (DGNs) are extensively employed in Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and their variants to approximate the data manifold, and data distribution on that manifold. However, training samples are often obtained based on preferences, costs, or convenience producing artifacts in the empirical data distribution e.g. the large fraction of smiling faces in the CelebA dataset or the large fraction of dark-haired individuals in FFHQ). {\em These inconsistencies will be reproduced when sampling from the trained DGN, which has far-reaching potential implications for fairness, data augmentation, anomaly detection, domain adaptation, and beyond.} In response, we develop a differential geometry based sampler -coined MaGNET- that, given any trained DGN, produces samples that are uniformly distributed on the learned manifold. We prove theoretically and empirically that our technique produces a uniform distribution on the manifold regardless of the training set distribution. We perform a range of experiments on various datasets and DGNs. One of them considers the state-of-the-art StyleGAN2 trained on FFHQ dataset, where uniform sampling via MaGNET increases distribution precision \& recall by 4.12\% \& 3.01\% and decreases gender bias by 41.2\%, without requiring labels or retraining.

Crowdsourcing has been instrumental for driving AI advances that rely on large-scale data. At the same time, reinforcement learning has seen rapid progress through benchmark environments that strike a balance between tractability and real-world complexity, such as ALE and OpenAI Gym. In this paper, we aim to fill a gap at the intersection of these two: The use of crowdsourcing to generate large-scale human demonstration data in the support of advancing research into imitation learning and offline learning. To this end, we present CrowdPlay, a complete crowdsourcing pipeline for any standard RL environment including OpenAI Gym (made available under an open-source license); a large-scale publicly available crowdsourced dataset of human gameplay demonstrations in Atari 2600 games, including multimodal behavior and human-human and human-AI multiagent data; offline learning benchmarks with extensive human data evaluation; and a detailed study of incentives, including real-time feedback to drive high quality data. We hope that this will drive the improvement in design of algorithms that account for the complexity of human, behavioral data and thereby enable a step forward in direction of effective learning for real-world settings. Our code and dataset are available at https://mgerstgrasser.github.io/crowdplay/.

In this article, we investigate the spectral behavior of random features kernel matrices of the type ${\bf K} = \mathbb{E}_{{\bf w}} \left[\sigma\left({\bf w}^{\sf T}{\bf x}_i\right)\sigma\left({\bf w}^{\sf T}{\bf x}_j\right)\right]_{i,j=1}^n$, with nonlinear function $\sigma(\cdot)$, data ${\bf x}_1, \ldots, {\bf x}_n \in \mathbb{R}^p$, and random projection vector ${\bf w} \in \mathbb{R}^p$ having i.i.d. entries. In a high-dimensional setting where the number of data $n$ and their dimension $p$ are both large and comparable, we show, under a Gaussian mixture model for the data, that the eigenspectrum of ${\bf K}$ is independent of the distribution of the i.i.d.(zero-mean and unit-variance) entries of ${\bf w}$, and only depends on $\sigma(\cdot)$ via its (generalized) Gaussian moments $\mathbb{E}_{z\sim \mathcal N(0,1)}[\sigma'(z)]$ and $\mathbb{E}_{z\sim \mathcal N(0,1)}[\sigma''(z)]$. As a result, for any kernel matrix ${\bf K}$ of the form above, we propose a novel random features technique, called Ternary Random Features (TRFs), that (i) asymptotically yields the same limiting kernel as the original ${\bf K}$ in a spectral sense and (ii) can be computed and stored much more efficiently, by wisely tuning (in a data-dependent manner) the function $\sigma$ and the random vector ${\bf w}$, both taking values in $\{-1,0,1\}$. The computation of the proposed random features requires no multiplication, and a factor of $b$ times less bits for storage compared to classical random features such as random Fourier features, with $b$ the number of bits to store full precision values. Besides, it appears in our experiments on real data that the substantial gains in computation and storage are accompanied with somewhat improved performances compared to state-of-the-art random features methods.

Self-driving cars must detect vehicles, pedestrians, and other traffic participants accurately to operate safely. Small, far-away, or highly occluded objects are particularly challenging because there is limited information in the LiDAR point clouds for detecting them. To address this challenge, we leverage valuable information from the past: in particular, data collected in past traversals of the same scene. We posit that these past data, which are typically discarded, provide rich contextual information for disambiguating the above-mentioned challenging cases. To this end, we propose a novel end-to-end trainable Hindsight framework to extract this contextual information from past traversals and store it in an easy-to-query data structure, which can then be leveraged to aid future 3D object detection of the same scene. We show that this framework is compatible with most modern 3D detection architectures and can substantially improve their average precision on multiple autonomous driving datasets, most notably by more than 300% on the challenging cases. Our code is available at https://github.com/YurongYou/Hindsight.

Partial-label learning (PLL) solves the multi-class classification problem, where each training instance is assigned a set of candidate labels that include the true label. Recent advances showed that PLL can be compatible with deep neural networks, which achieved state-of-the-art performance. However, most of the existing deep PLL methods focus on designing proper training objectives under various assumptions on the collected data, which may limit their performance when the collected data cannot satisfy the adopted assumptions. In this paper, we propose to exploit the learned intrinsic representation of the model to identify the true label in the training process, which does not rely on any assumptions on the collected data. We make two key contributions. As the first contribution, we empirically show that the class activation map (CAM), a simple technique for discriminating the learning patterns of each class in images, could surprisingly be utilized to make accurate predictions on selecting the true label from candidate labels. Unfortunately, as CAM is confined to image inputs with convolutional neural networks, we are yet unable to directly leverage CAM to address the PLL problem with general inputs and models. Thus, as the second contribution, we propose the class activation value (CAV), which owns similar properties of CAM, while CAV is versatile in various types of inputs and models. Building upon CAV, we propose a novel method named CAV Learning (CAVL) that selects the true label by the class with the maximum CAV for model training. Extensive experiments on various datasets demonstrate that our proposed CAVL method achieves state-of-the-art performance.

The architecture and the parameters of neural networks are often optimized independently, which requires costly retraining of the parameters whenever the architecture is modified. In this work we instead focus on growing the architecture without requiring costly retraining. We present a method that adds new neurons during training without impacting what is already learned, while improving the training dynamics. We achieve the latter by maximizing the gradients of the new weights and efficiently find the optimal initialization by means of the singular value decomposition (SVD). We call this technique Gradient Maximizing Growth (GradMax) and demonstrate its effectiveness in variety of vision tasks and architectures. We open sourced our code at https://github.com/google-research/growneuron

Self-training, a semi-supervised learning algorithm, leverages a large amount of unlabeled data to improve learning when the labeled data are limited. Despite empirical successes, its theoretical characterization remains elusive. To the best of our knowledge, this work establishes the first theoretical analysis for the known iterative self-training paradigm and formally proves the benefits of unlabeled data in both training convergence and generalization ability. To make our theoretical analysis feasible, we focus on the case of one-hidden-layer neural networks. However, theoretical understanding of iterative self-training is non-trivial even for a shallow neural network. One of the key challenges is that existing neural network landscape analysis built upon supervised learning no longer holds in the (semi-supervised) self-training paradigm. We address this challenge and prove that iterative self-training converges linearly with both convergence rate and generalization accuracy improved in the order of $1/\sqrt{M}$, where $M$ is the number of unlabeled samples. Extensive experiments from shallow neural networks to deep neural networks are also provided to justify the correctness of our established theoretical insights on self-training.

We explore a new perspective on video understanding by casting the video recognition problem as an image recognition task. Our approach rearranges input video frames into super images, which allow for training an image classifier directly to fulfill the task of action recognition, in exactly the same way as image classification. With such a simple idea, we show that transformer-based image classifiers alone can suffice for action recognition. In particular, our approach demonstrates strong and promising performance against SOTA methods on several public datasets including Kinetics400, Moments In Time, Something-Something V2 (SSV2), Jester and Diving48. We also experiment with the prevalent ResNet image classifiers in computer vision to further validate our idea. The results on both Kinetics400 and SSV2 are comparable to some of the best-performed CNN approaches based on spatio-temporal modeling. Our source codes and models are available at \url{https://github.com/IBM/sifar-pytorch}.