ICML 2021 Papers

Intersection over union (IoU) score, also named Jaccard Index, is one of the most fundamental evaluation methods in machine learning. The original IoU computation cannot provide non-zero gradients and thus cannot be directly optimized by nowadays deep learning methods. Several recent works generalized IoU for bounding box regression, but they are not straightforward to adapt for pixelwise prediction. In particular, the original IoU fails to provide effective gradients for the non-overlapping and location-deviation cases, which results in performance plateau. In this paper, we propose PixIoU, a generalized IoU for pixelwise prediction that is sensitive to the distance for non-overlapping cases and the locations in prediction. We provide proofs that PixIoU holds many nice properties as the original IoU. To optimize the PixIoU, we also propose a loss function that is proved to be submodular, hence we can apply the Lovász functions, the efficient surrogates for submodular functions for learning this loss. Experimental results show consistent performance improvements by learning PixIoU over the original IoU for several different pixelwise prediction tasks on Pascal VOC, VOT-2020 and Cityscapes.

We consider the problem of learning fair policies in (deep) cooperative multi-agent reinforcement learning (MARL). We formalize it in a principled way as the problem of optimizing a welfare function that explicitly encodes two important aspects of fairness: efficiency and equity. We provide a theoretical analysis of the convergence of policy gradient for this problem. As a solution method, we propose a novel neural network architecture, which is composed of two sub-networks specifically designed for taking into account these two aspects of fairness. In experiments, we demonstrate the importance of the two sub-networks for fair optimization. Our overall approach is general as it can accommodate any (sub)differentiable welfare function. Therefore, it is compatible with various notions of fairness that have been proposed in the literature (e.g., lexicographic maximin, generalized Gini social welfare function, proportional fairness). Our method is generic and can be implemented in various MARL settings: centralized training and decentralized execution, or fully decentralized. Finally, we experimentally validate our approach in various domains and show that it can perform much better than previous methods, both in terms of efficiency and equity.

In adversarial training (AT), the main focus has been the objective and optimizer while the model has been less studied, so that the models being used are still those classic ones in standard training (ST). Classic network architectures (NAs) are generally worse than searched NA in ST, which should be the same in AT. In this paper, we argue that NA and AT cannot be handled independently, since given a dataset, the optimal NA in ST would be no longer optimal in AT. That being said, AT is time-consuming itself; if we directly search NAs in AT over large search spaces, the computation will be practically infeasible. Thus, we propose diverse-structured network (DS-Net), to significantly reduce the size of the search space: instead of low-level operations, we only consider predefined atomic blocks, where an atomic block is a time-tested building block like the residual block. There are only a few atomic blocks and thus we can weight all atomic blocks rather than find the best one in a searched block of DS-Net, which is an essential tradeoff between exploring diverse structures and exploiting the best structures. Empirical results demonstrate the advantages of DS-Net, i.e., weighting the atomic blocks.

Training deep neural network models in the presence of corrupted supervision is challenging as the corrupted data points may significantly impact generalization performance. To alleviate this problem, we present an efficient robust algorithm that achieves strong guarantees without any assumption on the type of corruption and provides a unified framework for both classification and regression problems. Unlike many existing approaches that quantify the quality of the data points (e.g., based on their individual loss values), and filter them accordingly, the proposed algorithm focuses on controlling the collective impact of data points on the average gradient. Even when a corrupted data point failed to be excluded by our algorithm, the data point will have a very limited impact on the overall loss, as compared with state-of-the-art filtering methods based on loss values. Extensive experiments on multiple benchmark datasets have demonstrated the robustness of our algorithm under different types of corruption. Our code is available at \url{https://github.com/illidanlab/PRL}.

Learning curves model a classifier’s test error as a function of the number of training samples. Prior works show that learning curves can be used to select model parameters and extrapolate performance. We investigate how to use learning curves to evaluate design choices, such as pretraining, architecture, and data augmentation. We propose a method to robustly estimate learning curves, abstract their parameters into error and data-reliance, and evaluate the effectiveness of different parameterizations. Our experiments exemplify use of learning curves for analysis and yield several interesting observations.

Traditional supervised learning aims to train a classifier in the closed-set world, where training and test samples share the same label space. In this paper, we target a more challenging and re_x0002_alistic setting: open-set learning (OSL), where there exist test samples from the classes that are unseen during training. Although researchers have designed many methods from the algorith_x0002_mic perspectives, there are few methods that pro_x0002_vide generalization guarantees on their ability to achieve consistent performance on different train_x0002_ing samples drawn from the same distribution. Motivated by the transfer learning and probably approximate correct (PAC) theory, we make a bold attempt to study OSL by proving its general_x0002_ization error-given training samples with size n, the estimation error will get close to order Op(1/$\sqrt{}$n). This is the first study to provide a generalization bound for OSL, which we do by theoretically investigating the risk of the tar_x0002_get classifier on unknown classes. According to our theory, a novel algorithm, called auxiliary open-set risk (AOSR) is proposed to address the OSL problem. Experiments verify the efficacy of AOSR. The code is available at github.com/AnjinLiu/Openset_Learning_AOSR.

We address the problem of learning binary decision trees that partition data for some downstream task. We propose to learn discrete parameters (i.e., for tree traversals and node pruning) and continuous parameters (i.e., for tree split functions and prediction functions) simultaneously using argmin differentiation. We do so by sparsely relaxing a mixed-integer program for the discrete parameters, to allow gradients to pass through the program to continuous parameters. We derive customized algorithms to efficiently compute the forward and backward passes. This means that our tree learning procedure can be used as an (implicit) layer in arbitrary deep networks, and can be optimized with arbitrary loss functions. We demonstrate that our approach produces binary trees that are competitive with existing single tree and ensemble approaches, in both supervised and unsupervised settings. Further, apart from greedy approaches (which do not have competitive accuracies), our method is faster to train than all other tree-learning baselines we compare with.

Convex optimization with feedback is a framework where a learner relies on iterative queries and feedback to arrive at the minimizer of a convex function. The paradigm has gained significant popularity recently thanks to its scalability in large-scale optimization and machine learning. The repeated interactions, however, expose the learner to privacy risks from eavesdropping adversaries that observe the submitted queries. In this paper, we study how to optimally obfuscate the learner’s queries in convex optimization with first-order feedback, so that their learned optimal value is provably difficult to estimate for the eavesdropping adversary. We consider two formulations of learner privacy: a Bayesian formulation in which the convex function is drawn randomly, and a minimax formulation in which the function is fixed and the adversary’s probability of error is measured with respect to a minimax criterion. We show that, if the learner wants to ensure the probability of the adversary estimating accurately be kept below 1/L, then the overhead in query complexity is additive in L in the minimax formulation, but multiplicative in L in the Bayesian formulation. Compared to existing learner-private sequential learning models with binary feedback, our results apply to the significantly richer family of general convex functions with full-gradient feedback. Our proofs are largely enabled by tools from the theory of Dirichlet processes, as well as more sophisticated lines of analysis aimed at measuring the amount of information leakage under a full-gradient oracle.

Optimization of non-convex loss surfaces containing many local minima remains a critical problem in a variety of domains, including operations research, informatics, and material design. Yet, current techniques either require extremely high iteration counts or a large number of random restarts for good performance. In this work, we propose adapting recent developments in meta-learning to these many-minima problems by learning the optimization algorithm for various loss landscapes. We focus on problems from atomic structural optimization—finding low energy configurations of many-atom systems—including widely studied models such as bimetallic clusters and disordered silicon. We find that our optimizer learns a hopping behavior which enables efficient exploration and improves the rate of low energy minima discovery. Finally, our learned optimizers show promising generalization with efficiency gains on never before seen tasks (e.g. new elements or compositions). Code is available at https://learn2hop.page.link/github.

We propose a latent space energy-based prior model for text generation and classification. The model stands on a generator network that generates the text sequence based on a continuous latent vector. The energy term of the prior model couples a continuous latent vector and a symbolic one-hot vector, so that discrete category can be inferred from the observed example based on the continuous latent vector. Such a latent space coupling naturally enables incorporation of information bottleneck regularization to encourage the continuous latent vector to extract information from the observed example that is informative of the underlying category. In our learning method, the symbol-vector coupling, the generator network and the inference network are learned jointly. Our model can be learned in an unsupervised setting where no category labels are provided. It can also be learned in semi-supervised setting where category labels are provided for a subset of training examples. Our experiments demonstrate that the proposed model learns well-structured and meaningful latent space, which (1) guides the generator to generate text with high quality, diversity, and interpretability, and (2) effectively classifies text.

In this paper we present a scalable deep learning framework for finding Markovian Nash Equilibria in multi-agent stochastic games using fictitious play. The motivation is inspired by theoretical analysis of Forward Backward Stochastic Differential Equations and their implementation in a deep learning setting, which is the source of our algorithm’s sample efficiency improvement. By taking advantage of the permutation-invariant property of agents in symmetric games, the scalability and performance is further enhanced significantly. We showcase superior performance of our framework over the state-of-the-art deep fictitious play algorithm on an inter-bank lending/borrowing problem in terms of multiple metrics. More importantly, our approach scales up to 3000 agents in simulation, a scale which, to the best of our knowledge, represents a new state-of-the-art. We also demonstrate the applicability of our framework in robotics on a belief space autonomous racing problem.

Meta-learning of shared initialization parameters has shown to be highly effective in solving few-shot learning tasks. However, extending the framework to many-shot scenarios, which may further enhance its practicality, has been relatively overlooked due to the technical difficulties of meta-learning over long chains of inner-gradient steps. In this paper, we first show that allowing the meta-learners to take a larger number of inner gradient steps better captures the structure of heterogeneous and large-scale task distributions, thus results in obtaining better initialization points. Further, in order to increase the frequency of meta-updates even with the excessively long inner-optimization trajectories, we propose to estimate the required shift of the task-specific parameters with respect to the change of the initialization parameters. By doing so, we can arbitrarily increase the frequency of meta-updates and thus greatly improve the meta-level convergence as well as the quality of the learned initializations. We validate our method on a heterogeneous set of large-scale tasks, and show that the algorithm largely outperforms the previous first-order meta-learning methods in terms of both generalization performance and convergence, as well as multi-task learning and fine-tuning baselines.

\emph{Contrastive learning} (CL) pretrains models in a pairwise manner, where given a data point, other data points are all regarded as dissimilar, including some that are \emph{semantically} similar. The issue has been addressed by properly weighting similar and dissimilar pairs as in \emph{positive-unlabeled learning}, so that the objective of CL is \emph{unbiased} and CL is \emph{consistent}. However, in this paper, we argue that this great solution is still not enough: its weighted objective \emph{hides} the issue where the semantically similar pairs are still pushed away; as CL is pretraining, this phenomenon is not our desideratum and might affect downstream tasks. To this end, we propose \emph{large-margin contrastive learning} (LMCL) with \emph{distance polarization regularizer}, motivated by the distribution characteristic of pairwise distances in \emph{metric learning}. In LMCL, we can distinguish between \emph{intra-cluster} and \emph{inter-cluster} pairs, and then only push away inter-cluster pairs, which \emph{solves} the above issue explicitly. Theoretically, we prove a tighter error bound for LMCL; empirically, the superiority of LMCL is demonstrated across multiple domains, \emph{i.e.}, image classification, sentence representation, and reinforcement learning.

We propose a reparametrization scheme to address the challenges of applying differentially private SGD on large neural networks, which are 1) the huge memory cost of storing individual gradients, 2) the added noise suffering notorious dimensional dependence. Specifically, we reparametrize each weight matrix with two \emph{gradient-carrier} matrices of small dimension and a \emph{residual weight} matrix. We argue that such reparametrization keeps the forward/backward process unchanged while enabling us to compute the projected gradient without computing the gradient itself. To learn with differential privacy, we design \emph{reparametrized gradient perturbation (RGP)} that perturbs the gradients on gradient-carrier matrices and reconstructs an update for the original weight from the noisy gradients. Importantly, we use historical updates to find the gradient-carrier matrices, whose optimality is rigorously justified under linear regression and empirically verified with deep learning tasks. RGP significantly reduces the memory cost and improves the utility. For example, we are the first able to apply differential privacy on the BERT model and achieve an average accuracy of $83.9%$ on four downstream tasks with $\epsilon=8$, which is within $5%$ loss compared to the non-private baseline but enjoys much lower privacy leakage risk.

Membership inference is one of the simplest privacy threats faced by machine learning models that are trained on private sensitive data. In this attack, an adversary infers whether a particular point was used to train the model, or not, by observing the model’s predictions. Whereas current attack methods all require access to the model’s predicted confidence score, we introduce a label-only attack that instead evaluates the robustness of the model’s predicted (hard) labels under perturbations of the input, to infer membership. Our label-only attack is not only as-effective as attacks requiring access to confidence scores, it also demonstrates that a class of defenses against membership inference, which we call “confidence masking” because they obfuscate the confidence scores to thwart attacks, are insufficient to prevent the leakage of private information. Our experiments show that training with differential privacy or strong L2 regularization are the only current defenses that meaningfully decrease leakage of private information, even for points that are outliers of the training distribution.

We address the problem of teaching a deep reinforcement learning (RL) agent to follow instructions in multi-task environments. Instructions are expressed in a well-known formal language {–} linear temporal logic (LTL) {–} and can specify a diversity of complex, temporally extended behaviours, including conditionals and alternative realizations. Our proposed learning approach exploits the compositional syntax and the semantics of LTL, enabling our RL agent to learn task-conditioned policies that generalize to new instructions, not observed during training. To reduce the overhead of learning LTL semantics, we introduce an environment-agnostic LTL pretraining scheme which improves sample-efficiency in downstream environments. Experiments on discrete and continuous domains target combinatorial task sets of up to $\sim10^{39}$ unique tasks and demonstrate the strength of our approach in learning to solve (unseen) tasks, given LTL instructions.

While designing inductive bias in neural architectures has been widely studied, we hypothesize that transformer networks are flexible enough to learn inductive bias from suitable generic tasks. Here, we replace architecture engineering by encoding inductive bias in the form of datasets. Inspired by Peirce’s view that deduction, induction, and abduction are the primitives of reasoning, we design three synthetic tasks that are intended to require the model to have these three abilities. We specifically design these tasks to be synthetic and devoid of mathematical knowledge to ensure that only the fundamental reasoning biases can be learned from these tasks. This defines a new pre-training methodology called "LIME" (Learning Inductive bias for Mathematical rEasoning). Models trained with LIME significantly outperform vanilla transformers on four very different large mathematical reasoning benchmarks. Unlike dominating the computation cost as traditional pre-training approaches, LIME requires only a small fraction of the computation cost of the typical downstream task. The code for generating LIME tasks is available at https://github.com/tonywu95/LIME.

Answering complex natural language questions on knowledge graphs (KGQA) is a challenging task. It requires reasoning with the input natural language questions as well as a massive, incomplete heterogeneous KG. Prior methods obtain an abstract structured query graph/tree from the input question and traverse the KG for answers following the query tree. However, they inherently cannot deal with missing links in the KG. Here we present LEGO, a Latent Execution-Guided reasOning framework to handle this challenge in KGQA. LEGO works in an iterative way, which alternates between (1) a Query Synthesizer, which synthesizes a reasoning action and grows the query tree step-by-step, and (2) a Latent Space Executor that executes the reasoning action in the latent embedding space to combat against the missing information in KG. To learn the synthesizer without step-wise supervision, we design a generic latent execution guided bottom-up search procedure to find good execution traces efficiently in the vast query space. Experimental results on several KGQA benchmarks demonstrate the effectiveness of our framework compared with previous state of the art.

Face recognition is an important yet challenging problem in computer vision. A major challenge in practical face recognition applications lies in significant variations between profile and frontal faces. Traditional techniques address this challenge either by synthesizing frontal faces or by pose invariant learning. In this paper, we propose a novel method with Lie algebra theory to explore how face rotation in the 3D space affects the deep feature generation process of convolutional neural networks (CNNs). We prove that face rotation in the image space is equivalent to an additive residual component in the feature space of CNNs, which is determined solely by the rotation. Based on this theoretical finding, we further design a Lie Algebraic Residual Network (LARNet) for tackling pose robust face recognition. Our LARNet consists of a residual subnet for decoding rotation information from input face images, and a gating subnet to learn rotation magnitude for controlling the strength of the residual component contributing to the feature learning process. Comprehensive experimental evaluations on both frontal-profile face datasets and general face recognition datasets convincingly demonstrate that our method consistently outperforms the state-of-the-art ones.