ICLR 2022 Papers

Many computer vision systems require low-cost segmentation algorithms based on deep learning, either because of the enormous size of input images or limited computational budget. Common solutions uniformly downsample the input images to meet memory constraints, assuming all pixels are equally informative. In this work, we demonstrate that this assumption can harm the segmentation performance because the segmentation difficulty varies spatially (see Figure 1 “Uniform”). We combat this problem by introducing a learnable downsampling module, which can be optimised together with the given segmentation model in an end-to-end fashion. We formulate the problem of training such downsampling module as optimisation of sampling density distributions over the input images given their low-resolution views. To defend against degenerate solutions (e.g. over-sampling trivial regions like the backgrounds), we propose a regularisation term that encourages the sampling locations to concentrate around the object boundaries. We find the downsampling module learns to sample more densely at difficult locations, thereby improving the segmentation performance (see Figure 1 "Ours"). Our experiments on benchmarks of high-resolution street view, aerial and medical images demonstrate substantial improvements in terms of efficiency-and-accuracy trade-off compared to both uniform downsampling and two recent advanced downsampling techniques.

Recent work has shown that models trained to the same objective, and which achieve similar measures of accuracy on consistent test data, may nonetheless behave very differently on individual predictions. This inconsistency is undesirable in high-stakes contexts, such as medical diagnosis and finance. We show that this duplicitous behavior extends beyond predictions to feature attributions, which may likewise have negative implications for the intelligibility of a model, and one's ability to find recourse for subjects. We then introduce selective ensembles to mitigate such inconsistencies by applying hypothesis testing to the predictions of a set of models trained using randomly-selected starting conditions; importantly, selective ensembles can abstain in cases where a consistent outcome cannot be achieved up to a specified confidence level. We prove that that prediction disagreement between selective ensembles is bounded, and empirically demonstrate that selective ensembles achieve consistent predictions and feature attributions while maintaining low abstention rates. On several benchmark datasets, selective ensembles reach zero inconsistently predicted points, with abstention rates as low as 1.5%.

We propose Generative Pseudo-Inverse Memory (GPM), a class of deep generative memory models that are fast to write in and read out. Memory operations are recast as seeking robust solutions of linear systems, which naturally lead to the use of matrix pseudo-inverses. The pseudo-inverses are iteratively approximated, with practical computation complexity of almost $O(1)$. We prove theoretically and verify empirically that our model can retrieve exactly what have been written to the memory under mild conditions. A key capability of GPM is iterative reading, during which the attractor dynamics towards fixed points are enabled, allowing the model to iteratively improve sample quality in denoising and generating. More impressively, GPM can store a large amount of data while maintaining key abilities of accurate retrieving of stored patterns, denoising of corrupted data and generating novel samples. Empirically we demonstrate the efficiency and versatility of GPM on a comprehensive suite of experiments involving binarized MNIST, binarized Omniglot, FashionMNIST, CIFAR10 & CIFAR100 and CelebA.

The task of hot-refresh model upgrades of image retrieval systems plays an essential role in the industry but has never been investigated in academia before. Conventional cold-refresh model upgrades can only deploy new models after the gallery is overall backfilled, taking weeks or even months for massive data. In contrast, hot-refresh model upgrades deploy the new model immediately and then gradually improve the retrieval accuracy by backfilling the gallery on-the-fly. Compatible training has made it possible, however, the problem of model regression with negative flips poses a great challenge to the stable improvement of user experience. We argue that it is mainly due to the fact that new-to-old positive query-gallery pairs may show less similarity than new-to-new negative pairs. To solve the problem, we introduce a Regression-Alleviating Compatible Training (RACT) method to properly constrain the feature compatibility while reducing negative flips. The core is to encourage the new-to-old positive pairs to be more similar than both the new-to-old negative pairs and the new-to-new negative pairs. An efficient uncertainty-based backfilling strategy is further introduced to fasten accuracy improvements. Extensive experiments on large-scale retrieval benchmarks (e.g., Google Landmark) demonstrate that our RACT effectively alleviates the model regression for one more step towards seamless model upgrades.

Advanced representation learning techniques require reliable and general evaluation methods. Recently, several algorithms based on the common idea of geometric and topological analysis of a manifold approximated from the learned data representations have been proposed. In this work, we introduce Delaunay Component Analysis (DCA) -- an evaluation algorithm which approximates the data manifold using a more suitable neighbourhood graph called Delaunay graph. This provides a reliable manifold estimation even for challenging geometric arrangements of representations such as clusters with varying shape and density as well as outliers, which is where existing methods often fail. Furthermore, we exploit the nature of Delaunay graphs and introduce a framework for assessing the quality of individual novel data representations. We experimentally validate the proposed DCA method on representations obtained from neural networks trained with contrastive objective, supervised and generative models, and demonstrate various use cases of our extended single point evaluation framework.

As reinforcement learning (RL) has achieved great success and been even adopted in safety-critical domains such as autonomous vehicles, a range of empirical studies have been conducted to improve its robustness against adversarial attacks. However, how to certify its robustness with theoretical guarantees still remains challenging. In this paper, we present the first unified framework CROP (Certifying Robust Policies for RL) to provide robustness certification on both action and reward levels. In particular, we propose two robustness certification criteria: robustness of per-state actions and lower bound of cumulative rewards. We then develop a local smoothing algorithm for policies derived from Q-functions to guarantee the robustness of actions taken along the trajectory; we also develop a global smoothing algorithm for certifying the lower bound of a finite-horizon cumulative reward, as well as a novel local smoothing algorithm to perform adaptive search in order to obtain tighter reward certification. Empirically, we apply CROP to evaluate several existing empirically robust RL algorithms, including adversarial training and different robust regularization, in four environments (two representative Atari games, Highway, and CartPole). Furthermore, by evaluating these algorithms against adversarial attacks, we demonstrate that our certifications are often tight. All experiment results are available at website https://crop-leaderboard.github.io.

A key challenge in neural architecture search (NAS) is quickly inferring the predictive performance of a broad spectrum of networks to discover statistically accurate and computationally efficient ones. We refer to this task as model performance inference (MPI). The current practice for efficient MPI is gradient-based methods that leverage the gradients of a network at initialization to infer its performance. However, existing gradient-based methods rely only on heuristic metrics and lack the necessary theoretical foundations to consolidate their designs. We propose GradSign, an accurate, simple, and flexible metric for model performance inference with theoretical insights. The key idea behind GradSign is a quantity Ψ to analyze the sample-wise optimization landscape of different networks. Theoretically, we show that Ψ is an upper bound for both the training and true population losses of a neural network under reasonable assumptions. However, it is computationally prohibitive to directly calculate Ψ for modern neural networks. To address this challenge, we design GradSign, an accurate and simple approximation of Ψ using the gradients of a network evaluated at a random initialization state. Evaluation on seven NAS benchmarks across three training datasets shows that GradSign generalizes well to real-world networks and consistently outperforms state-of-the-art gradient-based methods for MPI evaluated by Spearman’s ρ and Kendall’s Tau. Additionally, we integrate GradSign into four existing NAS algorithms and show that the GradSign-assisted NAS algorithms outperform their vanilla counterparts by improving the accuracies of best-discovered networks by up to 0.3%, 1.1%, and 1.0% on three real-world tasks. Code is available at https://github.com/JackFram/GradSign

Q-learning is a popular Reinforcement Learning (RL) algorithm which is widely used in practice with function approximation (Mnih et al., 2015). In contrast, existing theoretical results are pessimistic about Q-learning. For example, (Baird, 1995) shows that Q-learning does not converge even with linear function approximation for linear MDPs. Furthermore, even for tabular MDPs with synchronous updates, Q-learning was shown to have sub-optimal sample complexity (Li et al., 2021, Azar et al., 2013). The goal of this work is to bridge the gap between practical success of Q-learning and the relatively pessimistic theoretical results. The starting point of our work is the observation that in practice, Q-learning is used with two important modifications: (i) training with two networks, called online network and target network simultaneously (online target learning, or OTL) , and (ii) experience replay (ER) (Mnih et al., 2015). While they have been observed to play a significant role in the practical success of Q-learning, a thorough theoretical understanding of how these two modifications improve the convergence behavior of Q-learning has been missing in literature. By carefully combining the Q-learning with OTL and reverse experience replay (RER) (a form of experience replay), we present novel methods Q-Rex and Q-RexDaRe (Q-Rex+data reuse). We show that Q-Rex efficiently finds the optimal policy for linear MDPs and provide non-asymptotic bounds on sample complexity -- the first such result for a Q-learning method with linear MDPs. Furthermore, we demonstrate that Q-RexDaRe in fact achieves near optimal sample complexity in the tabular setting, improving upon the existing results for vanilla Q-learning.

Self-supervised visual representation learning has recently attracted significant research interest. While a common way to evaluate self-supervised representations is through transfer to various downstream tasks, we instead investigate the problem of measuring their interpretability, i.e. understanding the semantics encoded in raw representations. We formulate the latter as estimating the mutual information between the representation and a space of manually labelled concepts. To quantify this we introduce a decoding bottleneck: information must be captured by simple predictors, mapping concepts to clusters in representation space. This approach, which we call reverse linear probing, provides a single number sensitive to the semanticity of the representation. This measure is also able to detect when the representation contains combinations of concepts (e.g., "red apple'') instead of just individual attributes ("red'' and "apple'' independently). Finally, we propose to use supervised classifiers to automatically label large datasets in order to enrich the space of concepts used for probing. We use our method to evaluate a large number of self-supervised representations, ranking them by interpretability, highlight the differences that emerge compared to the standard evaluation with linear probes and discuss several qualitative insights. Code at: https://github.com/iro-cp/ssl-qrp.

We present a novel adaptive optimization algorithm for large-scale machine learning problems. Equipped with a low-cost estimate of local curvature and Lipschitz smoothness, our method dynamically adapts the search direction and step-size. The search direction contains gradient information preconditioned by a well-scaled diagonal preconditioning matrix that captures the local curvature information. Our methodology does not require the tedious task of learning rate tuning, as the learning rate is updated automatically without adding an extra hyper-parameter. We provide convergence guarantees on a comprehensive collection of optimization problems, including convex, strongly convex, and nonconvex problems, in both deterministic and stochastic regimes. We also conduct an extensive empirical evaluation on standard machine learning problems, justifying our algorithm's versatility and demonstrating its strong performance compared to other start-of-the-art first-order and second-order methods.

Existing approaches to lifelong language learning rely on plenty of labeled data for learning a new task, which is hard to obtain in most real scenarios. Considering that humans can continually learn new tasks from a handful of examples, we expect the models also to be able to generalize well on new few-shot tasks without forgetting the previous ones. In this work, we define this more challenging yet practical problem as Lifelong Few-shot Language Learning (LFLL) and propose a unified framework for it based on prompt tuning of T5. Our framework called LFPT5 takes full advantage of PT's strong few-shot learning ability, and simultaneously trains the model as a task solver and a data generator. Before learning a new domain of the same task type, LFPT5 generates pseudo (labeled) samples of previously learned domains, and later gets trained on those samples to alleviate forgetting of previous knowledge as it learns the new domain. In addition, a KL divergence loss is minimized to achieve label consistency between the previous and the current model. While adapting to a new task type, LFPT5 includes and tunes additional prompt embeddings for the new task. With extensive experiments, we demonstrate that LFPT5 can be applied to various different types of tasks and significantly outperform previous methods in different LFLL settings.

Inferring interactions among entities is an important problem in studying dynamical systems, which greatly impacts the performance of downstream tasks, such as prediction. In this paper, we tackle the relational inference problem in a setting where each entity can potentially have a set of individualized information that other entities cannot have access to. Specifically, we represent the system using a graph in which the individualized information become node-specific information (NSI). We build our model in the framework of Neural Relation Inference (NRI), where the interaction among entities are uncovered using variational inference. We adopt NRI model to incorporate the individualized information by introducing private nodes in the graph that represent NSI. Such representation enables us to uncover more accurate relations among the agents and therefore leads to better performance on the downstream tasks. Our experiment results over real-world datasets validate the merit of our proposed algorithm.

This work targets automated designing and scaling of Vision Transformers (ViTs). The motivation comes from two pain spots: 1) the lack of efficient and principled methods for designing and scaling ViTs; 2) the tremendous computational cost of training ViT that is much heavier than its convolution counterpart. To tackle these issues, we propose As-ViT, an auto-scaling framework for ViTs without training, which automatically discovers and scales up ViTs in an efficient and principled manner. Specifically, we first design a "seed" ViT topology by leveraging a training-free search process. This extremely fast search is fulfilled by a comprehensive study of ViT's network complexity, yielding a strong Kendall-tau correlation with ground-truth accuracies. Second, starting from the "seed" topology, we automate the scaling rule for ViTs by growing widths/depths to different ViT layers. This results in a series of architectures with different numbers of parameters in a single run. Finally, based on the observation that ViTs can tolerate coarse tokenization in early training stages, we propose a progressive tokenization strategy to train ViTs faster and cheaper. As a unified framework, As-ViT achieves strong performance on classification (83.5% top1 on ImageNet-1k) and detection (52.7% mAP on COCO) without any manual crafting nor scaling of ViT architectures: the end-to-end model design and scaling process costs only 12 hours on one V100 GPU. Our code is available at https://github.com/VITA-Group/AsViT.

This paper introduces Multi-Agent MDP Homomorphic Networks, a class of networks that allows distributed execution using only local information, yet is able to share experience between global symmetries in the joint state-action space of cooperative multi-agent systems. In cooperative multi-agent systems, complex symmetries arise between different configurations of the agents and their local observations. For example, consider a group of agents navigating: rotating the state globally results in a permutation of the optimal joint policy. Existing work on symmetries in single agent reinforcement learning can only be generalized to the fully centralized setting, because such approaches rely on the global symmetry in the full state-action spaces, and these can result in correspondences across agents. To encode such symmetries while still allowing distributed execution we propose a factorization that decomposes global symmetries into local transformations. Our proposed factorization allows for distributing the computation that enforces global symmetries over local agents and local interactions. We introduce a multi-agent equivariant policy network based on this factorization. We show empirically on symmetric multi-agent problems that globally symmetric distributable policies improve data efficiency compared to non-equivariant baselines.

Rate-distortion (R-D) function, a key quantity in information theory, characterizes the fundamental limit of how much a data source can be compressed subject to a fidelity criterion, by any compression algorithm. As researchers push for ever-improving compression performance, establishing the R-D function of a given data source is not only of scientific interest, but also reveals the possible room for improvement in existing compression algorithms. Previous work on this problem relied on distributional assumptions on the data source (Gibson, 2017) or only applied to discrete data (Blahut, 1972; Arimoto, 1972). By contrast, this paper makes the first attempt at an algorithm for sandwiching the R-D function of a general (not necessarily discrete) source requiring only i.i.d. data samples. We estimate R-D sandwich bounds for a variety of artificial and real-world data sources, in settings far beyond the feasibility of any known method, and shed light on the optimality of neural data compression (Ballé et al., 2021; Yang et al., 2022). Our R-D upper bound on natural images indicates theoretical room for improving state-of-the-art image compression methods by at least one dB in PSNR at various bitrates. Our data and code can be found at https://github.com/mandt-lab/RD-sandwich.

This paper considers few-shot learning under the cross-domain scenario. The cross-domain setting imposes a critical challenge, i.e., using very few (support) samples to generalize the already-learned model to a novel domain. We hold a hypothesis, i.e., if a deep model is capable to fast generalize itself to different domains (using very few samples) during training, it will maintain such domain generalization capacity for testing. It motivates us to propose a novel Domain-Switch Learning (DSL) framework. DSL embeds the cross-domain scenario into the training stage in a ``fast switching'' manner. Specifically, DSL uses a single domain for a training iteration and switches into another domain for the following iteration. During the switching, DSL enforces two constraints: 1) the deep model should not over-fit the domain in the current iteration and 2) the deep model should not forget the already-learned knowledge of other domains. These two constraints jointly promote fast generalization across different domains. Experimental results confirm that the cross-domain generalization capacity can be inherited from the training stage to the testing stage, validating our key hypothesis. Consequentially, DSL significantly improves cross-domain few-shot classification and sets up new state of the art.

In this work we study the MARINA method of Gorbunov et al (ICML, 2021) -- the current state-of-the-art distributed non-convex optimization method in terms of theoretical communication complexity. Theoretical superiority of this method can be largely attributed to two sources: a carefully engineered biased stochastic gradient estimator, which leads to a reduction in the number of communication rounds, and the reliance on {\em independent} stochastic communication compression, which leads to a reduction in the number of transmitted bits within each communication round. In this paper we i) extend the theory of MARINA to support a much wider class of potentially {\em correlated} compressors, extending the reach of the method beyond the classical independent compressors setting, ii) show that a new quantity, for which we coin the name {\em Hessian variance}, allows us to significantly refine the original analysis of MARINA without any additional assumptions, and iii) identify a special class of correlated compressors based on the idea of {\em random permutations}, for which we coin the term Perm$K$, the use of which leads to up to $O(\sqrt{n})$ (resp. $O(1 + d/\sqrt{n})$) improvement in the theoretical communication complexity of MARINA in the low Hessian variance regime when $d\geq n$ (resp. $d \leq n$), where $n$ is the number of workers and $d$ is the number of parameters describing the model we are learning. We corroborate our theoretical results with carefully engineered synthetic experiments with minimizing the average of nonconvex quadratics, and on autoencoder training with the MNIST dataset.

To attain higher efficiency, the industry has gradually reformed towards application-specific hardware accelerators. While such a paradigm shift is already starting to show promising results, designers need to spend considerable manual effort and perform large number of time-consuming simulations to find accelerators that can accelerate multiple target applications while obeying design constraints. Moreover, such a simulation-driven approach must be re-run from scratch every time the set of target applications or design constraints change. An alternative paradigm is to use a data-driven, offline approach that utilizes logged simulation data, to architect hardware accelerators, without needing any form of simulations. Such an approach not only alleviates the need to run time-consuming simulation, but also enables data reuse and applies even when set of target applications changes. In this paper, we develop such a data-driven offline optimization method for designing hardware accelerators, dubbed PRIME, that enjoys all of these properties. Our approach learns a conservative, robust estimate of the desired cost function, utilizes infeasible points and optimizes the design against this estimate without any additional simulator queries during optimization. PRIME architects accelerators---tailored towards both single- and multi-applications---improving performance upon stat-of-the-art simulation-driven methods by about 1.54x and 1.20x, while considerably reducing the required total simulation time by 93% and 99%, respectively. In addition, PRIME also architects effective accelerators for unseen applications in a zero-shot setting, outperforming simulation-based methods by 1.26x.

Catastrophic forgetting presents a challenge in developing deep learning models capable of continual learning, i.e. learning tasks sequentially. Recently, both computer vision and natural-language processing have witnessed great progress through the use of large-scale pretrained models. In this work, we present an empirical study of catastrophic forgetting in this pretraining paradigm. Our experiments indicate that large, pretrained ResNets and Transformers are significantly more resistant to forgetting than randomly-initialized, trained-from-scratch models; this robustness systematically improves with scale of both model and pretraining dataset size. We take initial steps towards characterizing what aspect of model representations allows them to perform continual learning so well, finding that in the pretrained models, distinct class representations grow more orthogonal with scale. Our results suggest that, when possible, scale and a diverse pretraining dataset can be useful ingredients in mitigating catastrophic forgetting.