ICML 2021 Papers

Solving sparse reward tasks through exploration is one of the major challenges in deep reinforcement learning, especially in three-dimensional, partially-observable environments. Critically, the algorithm proposed in this article is capable of using a single human demonstration to solve hard-exploration problems. We train an agent on a combination of demonstrations and own experience to solve problems with variable initial conditions and we integrate it with proximal policy optimization (PPO). The agent is also able to increase its performance and to tackle harder problems by replaying its own past trajectories prioritizing them based on the obtained reward and the maximum value of the trajectory. We finally compare variations of this algorithm to different imitation learning algorithms on a set of hard-exploration tasks in the Animal-AI Olympics environment. To the best of our knowledge, learning a task in a three-dimensional environment with comparable difficulty has never been considered before using only one human demonstration.

Choosing the right parameters for optimization algorithms is often the key to their success in practice. Solving this problem using a learning-to-learn approach—using meta-gradient descent on a meta-objective based on the trajectory that the optimizer generates—was recently shown to be effective. However, the meta-optimization problem is difficult. In particular, the meta-gradient can often explode/vanish, and the learned optimizer may not have good generalization performance if the meta-objective is not chosen carefully. In this paper we give meta-optimization guarantees for the learning-to-learn approach on a simple problem of tuning the step size for quadratic loss. Our results show that the naïve objective suffers from meta-gradient explosion/vanishing problem. Although there is a way to design the meta-objective so that the meta-gradient remains polynomially bounded, computing the meta-gradient directly using backpropagation leads to numerical issues. We also characterize when it is necessary to compute the meta-objective on a separate validation set to ensure the generalization performance of the learned optimizer. Finally, we verify our results empirically and show that a similar phenomenon appears even for more complicated learned optimizers parametrized by neural networks.

Sparse regression has recently been applied to enable transfer learning from very limited data. We study an extension of this approach to unsupervised learning—in particular, learning word embeddings from unstructured text corpora using low-rank matrix factorization. Intuitively, when transferring word embeddings to a new domain, we expect that the embeddings change for only a small number of words—e.g., the ones with novel meanings in that domain. We propose a novel group-sparse penalty that exploits this sparsity to perform transfer learning when there is very little text data available in the target domain—e.g., a single article of text. We prove generalization bounds for our algorithm. Furthermore, we empirically evaluate its effectiveness, both in terms of prediction accuracy in downstream tasks as well as in terms of interpretability of the results.

Network compression has been widely studied since it is able to reduce the memory and computation cost during inference. However, previous methods seldom deal with complicated structures like residual connections, group/depth-wise convolution and feature pyramid network, where channels of multiple layers are coupled and need to be pruned simultaneously. In this paper, we present a general channel pruning approach that can be applied to various complicated structures. Particularly, we propose a layer grouping algorithm to find coupled channels automatically. Then we derive a unified metric based on Fisher information to evaluate the importance of a single channel and coupled channels. Moreover, we find that inference speedup on GPUs is more correlated with the reduction of memory rather than FLOPs, and thus we employ the memory reduction of each channel to normalize the importance. Our method can be used to prune any structures including those with coupled channels. We conduct extensive experiments on various backbones, including the classic ResNet and ResNeXt, mobile-friendly MobileNetV2, and the NAS-based RegNet, both on image classification and object detection which is under-explored. Experimental results validate that our method can effectively prune sophisticated networks, boosting inference speed without sacrificing accuracy.

We investigate the use of natural language to drive the generalization of control policies and introduce the new multi-task environment Messenger with free-form text manuals describing the environment dynamics. Unlike previous work, Messenger does not assume prior knowledge connecting text and state observations {—} the control policy must simultaneously ground the game manual to entity symbols and dynamics in the environment. We develop a new model, EMMA (Entity Mapper with Multi-modal Attention) which uses an entity-conditioned attention module that allows for selective focus over relevant descriptions in the manual for each entity in the environment. EMMA is end-to-end differentiable and learns a latent grounding of entities and dynamics from text to observations using only environment rewards. EMMA achieves successful zero-shot generalization to unseen games with new dynamics, obtaining a 40% higher win rate compared to multiple baselines. However, win rate on the hardest stage of Messenger remains low (10%), demonstrating the need for additional work in this direction.

We introduce a new neural network architecture that we call "grid-functioned" neural networks. It utilises a grid structure of network parameterisations that can be specialised for different subdomains of the problem, while maintaining smooth, continuous behaviour. The grid gives the user flexibility to prevent gross features from overshadowing important minor ones. We present a full characterisation of its computational and spatial complexity, and demonstrate its potential, compared to a traditional architecture, over a set of synthetic regression problems. We further illustrate the benefits through a real-world 3D skeletal animation case study, where it offers the same visual quality as a state-of-the-art model, but with lower computational complexity and better control accuracy.

Model extraction attacks attempt to replicate a target machine learning model by querying its inference API. State-of-the-art attacks are learning-based and construct replicas by supervised training on the target model’s predictions, but an emerging class of attacks exploit algebraic properties to obtain high-fidelity replicas using orders of magnitude fewer queries. So far, these algebraic attacks have been limited to neural networks with few hidden layers and ReLU activations. In this paper we present algebraic and hybrid algebraic/learning-based attacks on large-scale natural language models. We consider a grey-box setting, targeting models with a pre-trained (public) encoder followed by a single (private) classification layer. Our key findings are that (i) with a frozen encoder, high-fidelity extraction is possible with a small number of in-distribution queries, making extraction attacks indistinguishable from legitimate use; (ii) when the encoder is fine-tuned, a hybrid learning-based/algebraic attack improves over the learning-based state-of-the-art without requiring additional queries.

Normalization is known to help the optimization of deep neural networks. Curiously, different architectures require specialized normalization methods. In this paper, we study what normalization is effective for Graph Neural Networks (GNNs). First, we adapt and evaluate the existing methods from other domains to GNNs. Faster convergence is achieved with InstanceNorm compared to BatchNorm and LayerNorm. We provide an explanation by showing that InstanceNorm serves as a preconditioner for GNNs, but such preconditioning effect is weaker with BatchNorm due to the heavy batch noise in graph datasets. Second, we show that the shift operation in InstanceNorm results in an expressiveness degradation of GNNs for highly regular graphs. We address this issue by proposing GraphNorm with a learnable shift. Empirically, GNNs with GraphNorm converge faster compared to GNNs using other normalization. GraphNorm also improves the generalization of GNNs, achieving better performance on graph classification benchmarks.

We consider the problem of molecular graph generation using deep models. While graphs are discrete, most existing methods use continuous latent variables, resulting in inaccurate modeling of discrete graph structures. In this work, we propose GraphDF, a novel discrete latent variable model for molecular graph generation based on normalizing flow methods. GraphDF uses invertible modulo shift transforms to map discrete latent variables to graph nodes and edges. We show that the use of discrete latent variables reduces computational costs and eliminates the negative effect of dequantization. Comprehensive experimental results show that GraphDF outperforms prior methods on random generation, property optimization, and constrained optimization tasks.

Despite the recent success of graph neural networks (GNN), common architectures often exhibit significant limitations, including sensitivity to oversmoothing, long-range dependencies, and spurious edges, e.g., as can occur as a result of graph heterophily or adversarial attacks. To at least partially address these issues within a simple transparent framework, we consider a new family of GNN layers designed to mimic and integrate the update rules of two classical iterative algorithms, namely, proximal gradient descent and iterative reweighted least squares (IRLS). The former defines an extensible base GNN architecture that is immune to oversmoothing while nonetheless capturing long-range dependencies by allowing arbitrary propagation steps. In contrast, the latter produces a novel attention mechanism that is explicitly anchored to an underlying end-to-end energy function, contributing stability with respect to edge uncertainty. When combined we obtain an extremely simple yet robust model that we evaluate across disparate scenarios including standardized benchmarks, adversarially-perturbated graphs, graphs with heterophily, and graphs involving long-range dependencies. In doing so, we compare against SOTA GNN approaches that have been explicitly designed for the respective task, achieving competitive or superior node classification accuracy. Our code is available at https://github.com/FFTYYY/TWIRLS. And for an extended version of this work, please see https://arxiv.org/abs/2103.06064.

We introduce the Graph Mixture Density Networks, a new family of machine learning models that can fit multimodal output distributions conditioned on graphs of arbitrary topology. By combining ideas from mixture models and graph representation learning, we address a broader class of challenging conditional density estimation problems that rely on structured data. In this respect, we evaluate our method on a new benchmark application that leverages random graphs for stochastic epidemic simulations. We show a significant improvement in the likelihood of epidemic outcomes when taking into account both multimodality and structure. The empirical analysis is complemented by two real-world regression tasks showing the effectiveness of our approach in modeling the output prediction uncertainty. Graph Mixture Density Networks open appealing research opportunities in the study of structure-dependent phenomena that exhibit non-trivial conditional output distributions.

We prove that the alpha-expansion algorithm for MAP inference always returns a globally optimal assignment for Markov Random Fields with Potts pairwise potentials, with a catch: the returned assignment is only guaranteed to be optimal for an instance within a small perturbation of the original problem instance. In other words, all local minima with respect to expansion moves are global minima to slightly perturbed versions of the problem. On "real-world" instances, MAP assignments of small perturbations of the problem should be very similar to the MAP assignment(s) of the original problem instance. We design an algorithm that can certify whether this is the case in practice. On several MAP inference problem instances from computer vision, this algorithm certifies that MAP solutions to all of these perturbations are very close to solutions of the original instance. These results taken together give a cohesive explanation for the good performance of "graph cuts" algorithms in practice. Every local expansion minimum is a global minimum in a small perturbation of the problem, and all of these global minima are close to the original solution.

Recently there has been increased interest in semi-supervised classification in the presence of graphical information. A new class of learning models has emerged that relies, at its most basic level, on classifying the data after first applying a graph convolution. To understand the merits of this approach, we study the classification of a mixture of Gaussians, where the data corresponds to the node attributes of a stochastic block model. We show that graph convolution extends the regime in which the data is linearly separable by a factor of roughly $1/\sqrt{D}$, where $D$ is the expected degree of a node, as compared to the mixture model data on its own. Furthermore, we find that the linear classifier obtained by minimizing the cross-entropy loss after the graph convolution generalizes to out-of-distribution data where the unseen data can have different intra- and inter-class edge probabilities from the training data.

Self-supervised learning on graph-structured data has drawn recent interest for learning generalizable, transferable and robust representations from unlabeled graphs. Among many, graph contrastive learning (GraphCL) has emerged with promising representation learning performance. Unfortunately, unlike its counterpart on image data, the effectiveness of GraphCL hinges on ad-hoc data augmentations, which have to be manually picked per dataset, by either rules of thumb or trial-and-errors, owing to the diverse nature of graph data. That significantly limits the more general applicability of GraphCL. Aiming to fill in this crucial gap, this paper proposes a unified bi-level optimization framework to automatically, adaptively and dynamically select data augmentations when performing GraphCL on specific graph data. The general framework, dubbed JOint Augmentation Optimization (JOAO), is instantiated as min-max optimization. The selections of augmentations made by JOAO are shown to be in general aligned with previous "best practices" observed from handcrafted tuning: yet now being automated, more flexible and versatile. Moreover, we propose a new augmentation-aware projection head mechanism, which will route output features through different projection heads corresponding to different augmentations chosen at each training step. Extensive experiments demonstrate that JOAO performs on par with or sometimes better than the state-of-the-art competitors including GraphCL, on multiple graph datasets of various scales and types, yet without resorting to any laborious dataset-specific tuning on augmentation selection. We release the code at https://github.com/Shen-Lab/GraphCL_Automated.

We show that aggregated model updates in federated learning may be insecure. An untrusted central server may disaggregate user updates from sums of updates across participants given repeated observations, enabling the server to recover privileged information about individual users’ private training data via traditional gradient inference attacks. Our method revolves around reconstructing participant information (e.g: which rounds of training users participated in) from aggregated model updates by leveraging summary information from device analytics commonly used to monitor, debug, and manage federated learning systems. Our attack is parallelizable and we successfully disaggregate user updates on settings with up to thousands of participants. We quantitatively and qualitatively demonstrate significant improvements in the capability of various inference attacks on the disaggregated updates. Our attack enables the attribution of learned properties to individual users, violating anonymity, and shows that a determined central server may undermine the secure aggregation protocol to break individual users’ data privacy in federated learning.

Recently, denoising diffusion probabilistic models and generative score matching have shown high potential in modelling complex data distributions while stochastic calculus has provided a unified point of view on these techniques allowing for flexible inference schemes. In this paper we introduce Grad-TTS, a novel text-to-speech model with score-based decoder producing mel-spectrograms by gradually transforming noise predicted by encoder and aligned with text input by means of Monotonic Alignment Search. The framework of stochastic differential equations helps us to generalize conventional diffusion probabilistic models to the case of reconstructing data from noise with different parameters and allows to make this reconstruction flexible by explicitly controlling trade-off between sound quality and inference speed. Subjective human evaluation shows that Grad-TTS is competitive with state-of-the-art text-to-speech approaches in terms of Mean Opinion Score.

Goal-conditioned reinforcement learning endows an agent with a large variety of skills, but it often struggles to solve tasks that require more temporally extended reasoning. In this work, we propose to incorporate imagined subgoals into policy learning to facilitate learning of complex tasks. Imagined subgoals are predicted by a separate high-level policy, which is trained simultaneously with the policy and its critic. This high-level policy predicts intermediate states halfway to the goal using the value function as a reachability metric. We don’t require the policy to reach these subgoals explicitly. Instead, we use them to define a prior policy, and incorporate this prior into a KL-constrained policy iteration scheme to speed up and regularize learning. Imagined subgoals are used during policy learning, but not during test time, where we only apply the learned policy. We evaluate our approach on complex robotic navigation and manipulation tasks and show that it outperforms existing methods by a large margin.

The threat of adversarial examples has motivated work on training certifiably robust neural networks to facilitate efficient verification of local robustness at inference time. We formalize a notion of global robustness, which captures the operational properties of on-line local robustness certification while yielding a natural learning objective for robust training. We show that widely-used architectures can be easily adapted to this objective by incorporating efficient global Lipschitz bounds into the network, yielding certifiably-robust models by construction that achieve state-of-the-art verifiable accuracy. Notably, this approach requires significantly less time and memory than recent certifiable training methods, and leads to negligible costs when certifying points on-line; for example, our evaluation shows that it is possible to train a large robust Tiny-Imagenet model in a matter of hours. Our models effectively leverage inexpensive global Lipschitz bounds for real-time certification, despite prior suggestions that tighter local bounds are needed for good performance; we posit this is possible because our models are specifically trained to achieve tighter global bounds. Namely, we prove that the maximum achievable verifiable accuracy for a given dataset is not improved by using a local bound.

We consider the optimal approximate posterior over the top-layer weights in a Bayesian neural network for regression, and show that it exhibits strong dependencies on the lower-layer weights. We adapt this result to develop a correlated approximate posterior over the weights at all layers in a Bayesian neural network. We extend this approach to deep Gaussian processes, unifying inference in the two model classes. Our approximate posterior uses learned "global” inducing points, which are defined only at the input layer and propagated through the network to obtain inducing inputs at subsequent layers. By contrast, standard, "local”, inducing point methods from the deep Gaussian process literature optimise a separate set of inducing inputs at every layer, and thus do not model correlations across layers. Our method gives state-of-the-art performance for a variational Bayesian method, without data augmentation or tempering, on CIFAR-10 of 86.7%, which is comparable to SGMCMC without tempering but with data augmentation (88% in Wenzel et al. 2020).