NeurIPS 2021 Papers

Computational methods that operate on three-dimensional (3D) molecular structure have the potential to solve important problems in biology and chemistry. Deep neural networks have gained significant attention, but their widespread adoption in the biomolecular domain has been limited by a lack of either systematic performance benchmarks or a unified toolkit for interacting with 3D molecular data. To address this, we present ATOM3D, a collection of both novel and existing benchmark datasets spanning several key classes of biomolecules. We implement several types of 3D molecular learning methods for each of these tasks and show that they consistently improve performance relative to methods based on one- and two-dimensional representations. The choice of architecture proves to be important for performance, with 3D convolutional networks excelling at tasks involving complex geometries, graph networks performing well on systems requiring detailed positional information, and the more recently developed equivariant networks showing significant promise. Our results indicate that many molecular problems stand to gain from 3D molecular learning, and that there is potential for substantial further improvement on many tasks. To lower the barrier to entry and facilitate further developments in the field, we also provide a comprehensive suite of tools for dataset processing, model training, and evaluation in our open-source atom3d Python package. All datasets are available for download from www.atom3d.ai.

We propose a framework for comparing data manifolds, aimed, in particular, towards the evaluation of deep generative models. We describe a novel tool, Cross-Barcode(P,Q), that, given a pair of distributions in a high-dimensional space, tracks multiscale topology spacial discrepancies between manifolds on which the distributions are concentrated. Based on the Cross-Barcode, we introduce the Manifold Topology Divergence score (MTop-Divergence) and apply it to assess the performance of deep generative models in various domains: images, 3D-shapes, time-series, and on different datasets: MNIST, Fashion MNIST, SVHN, CIFAR10, FFHQ, market stock data, ShapeNet. We demonstrate that the MTop-Divergence accurately detects various degrees of mode-dropping, intra-mode collapse, mode invention, and image disturbance. Our algorithm scales well (essentially linearly) with the increase of the dimension of the ambient high-dimensional space. It is one of the first TDA-based methodologies that can be applied universally to datasets of different sizes and dimensions, including the ones on which the most recent GANs in the visual domain are trained. The proposed method is domain agnostic and does not rely on pre-trained networks.

The automatic generation of long and coherent medical reports given medical images (e.g. Chest X-ray and Fundus Fluorescein Angiography (FFA)) has great potential to support clinical practice. Researchers have explored advanced methods from computer vision and natural language processing to incorporate medical domain knowledge for the generation of readable medical reports. However, existing medical report generation (MRG) benchmarks lack both explainable annotations and reliable evaluation tools, hindering the current research advances from two aspects: firstly, existing methods can only predict reports without accurate explanation, undermining the trustworthiness of the diagnostic methods; secondly, the comparison among the predicted reports from different MRG methods is unreliable using the evaluation metrics of natural-language generation (NLG). To address these issues, in this paper, we propose an explainable and reliable MRG benchmark based on FFA Images and Reports (FFA-IR). Specifically, FFA-IR is large, with 10,790 reports along with 1,048,584 FFA images from clinical practice; it includes explainable annotations, based on a schema of 46 categories of lesions; and it is bilingual, providing both English and Chinese reports for each case. Besides using the widely used NLG metrics, we propose a set of nine human evaluation criteria to evaluate the generated reports. We envision FFA-IR as a testbed for explainable and reliable medical report generation. We also hope that it can broadly accelerate medical imaging research and facilitate interaction between the fields of medical imaging, computer vision, and natural language processing.

Meta-reinforcement learning (meta-RL) has proven to be a successful framework for leveraging experience from prior tasks to rapidly learn new related tasks, however, current meta-RL approaches struggle to learn in sparse reward environments. Although existing meta-RL algorithms can learn strategies for adapting to new sparse reward tasks, the actual adaptation strategies are learned using hand-shaped reward functions, or require simple environments where random exploration is sufficient to encounter sparse reward. In this paper we present a formulation of hindsight relabelling for meta-RL, which relabels experience during meta-training to enable learning to learn entirely using sparse reward. We demonstrate the effectiveness of our approach on a suite of challenging sparse reward environments that previously required dense reward during meta-training to solve. Our approach solves these environments using the true sparse reward function, with performance comparable to training with a proxy dense reward function.

Probabilistic numerics casts numerical tasks, such the numerical solution of differential equations, as inference problems to be solved. One approach is to model the unknown quantity of interest as a random variable, and to constrain this variable using data generated during the course of a traditional numerical method. However, data may be nonlinearly related to the quantity of interest, rendering the proper conditioning of random variables difficult and limiting the range of numerical tasks that can be addressed. Instead, this paper proposes to construct probabilistic numerical methods based only on the final output from a traditional method. A convergent sequence of approximations to the quantity of interest constitute a dataset, from which the limiting quantity of interest can be extrapolated, in a probabilistic analogue of Richardson’s deferred approach to the limit. This black box approach (1) massively expands the range of tasks to which probabilistic numerics can be applied, (2) inherits the features and performance of state-of-the-art numerical methods, and (3) enables provably higher orders of convergence to be achieved. Applications are presented for nonlinear ordinary and partial differential equations, as well as for eigenvalue problems—a setting for which no probabilistic numerical methods have yet been developed.

Numerical simulations of Earth's weather and climate require substantial amounts of computation. This has led to a growing interest in replacing subroutines that explicitly compute physical processes with approximate machine learning (ML) methods that are fast at inference time. Within weather and climate models, atmospheric radiative transfer (RT) calculations are especially expensive. This has made them a popular target for neural network-based emulators. However, prior work is hard to compare due to the lack of a comprehensive dataset and standardized best practices for ML benchmarking. To fill this gap, we build a large dataset, ClimART, with more than 10 million samples from present, pre-industrial, and future climate conditions, based on the Canadian Earth System Model. ClimART poses several methodological challenges for the ML community, such as multiple out-of-distribution test sets, underlying domain physics, and a trade-off between accuracy and inference speed. We also present several novel baselines that indicate shortcomings of datasets and network architectures used in prior work.

Modern deep learning applications require increasingly more compute to train state-of-the-art models. To address this demand, large corporations and institutions use dedicated High-Performance Computing clusters, whose construction and maintenance are both environmentally costly and well beyond the budget of most organizations. As a result, some research directions become the exclusive domain of a few large industrial and even fewer academic actors. To alleviate this disparity, smaller groups may pool their computational resources and run collaborative experiments that benefit all participants. This paradigm, known as grid- or volunteer computing, has seen successful applications in numerous scientific areas. However, using this approach for machine learning is difficult due to high latency, asymmetric bandwidth, and several challenges unique to volunteer computing. In this work, we carefully analyze these constraints and propose a novel algorithmic framework designed specifically for collaborative training. We demonstrate the effectiveness of our approach for SwAV and ALBERT pretraining in realistic conditions and achieve performance comparable to traditional setups at a fraction of the cost. Finally, we provide a detailed report of successful collaborative language model pretraining with nearly 50 participants.

While attention is all you need may be proving true, we do not know why: attention-based transformer models such as BERT are superior but how information flows from input tokens to output predictions are unclear. We introduce influence patterns, abstractions of sets of paths through a transformer model. Patterns quantify and localize the flow of information to paths passing through a sequence of model nodes. Experimentally, we find that significant portion of information flow in BERT goes through skip connections instead of attention heads. We further show that consistency of patterns across instances is an indicator of BERT’s performance. Finally, we demonstrate that patterns account for far more model performance than previous attention-based and layer-based methods.

Augmenting pre-trained language models with knowledge graphs (KGs) has achieved success on various commonsense reasoning tasks. However, for a given task instance, the KG, or certain parts of the KG, may not be useful. Although KG-augmented models often use attention to focus on specific KG components, the KG is still always used, and the attention mechanism is never explicitly taught which KG components should be used. Meanwhile, saliency methods can measure how much a KG feature (e.g., graph, node, path) influences the model to make the correct prediction, thus explaining which KG features are useful. This paper explores how saliency explanations can be used to improve KG-augmented models' performance. First, we propose to create coarse (Is the KG useful?) and fine (Which nodes/paths in the KG are useful?) saliency explanations. Second, to motivate saliency-based supervision, we analyze oracle KG-augmented models which directly use saliency explanations as extra inputs for guiding their attention. Third, we propose SalKG, a framework for KG-augmented models to learn from coarse and/or fine saliency explanations. Given saliency explanations created from a task's training set, SalKG jointly trains the model to predict the explanations, then solve the task by attending to KG features highlighted by the predicted explanations. On three popular commonsense QA benchmarks (CSQA, OBQA, CODAH) and a range of KG-augmented models, we show that SalKG can yield considerable performance gains --- up to 2.76% absolute improvement on CSQA.

Certified robustness is a desirable property for deep neural networks in safety-critical applications, and popular training algorithms can certify robustness of a neural network by computing a global bound on its Lipschitz constant. However, such a bound is often loose: it tends to over-regularize the neural network and degrade its natural accuracy. A tighter Lipschitz bound may provide a better tradeoff between natural and certified accuracy, but is generally hard to compute exactly due to non-convexity of the network. In this work, we propose an efficient and trainable \emph{local} Lipschitz upper bound by considering the interactions between activation functions (e.g. ReLU) and weight matrices. Specifically, when computing the induced norm of a weight matrix, we eliminate the corresponding rows and columns where the activation function is guaranteed to be a constant in the neighborhood of each given data point, which provides a provably tighter bound than the global Lipschitz constant of the neural network. Our method can be used as a plug-in module to tighten the Lipschitz bound in many certifiable training algorithms. Furthermore, we propose to clip activation functions (e.g., ReLU and MaxMin) with a learnable upper threshold and a sparsity loss to assist the network to achieve an even tighter local Lipschitz bound. Experimentally, we show that our method consistently outperforms state-of-the-art methods in both clean and certified accuracy on MNIST, CIFAR-10 and TinyImageNet datasets with various network architectures.

Most existing set encoding algorithms operate under the implicit assumption that all the set elements are accessible, and that there are ample computational and memory resources to load the set into memory during training and inference. However, both assumptions fail when the set is excessively large such that it is impossible to load all set elements into memory, or when data arrives in a stream. To tackle such practical challenges in large-scale set encoding, the general set-function constraints of permutation invariance and equivariance are not sufficient. We introduce a new property termed Mini-Batch Consistency (MBC) that is required for large scale mini-batch set encoding. Additionally, we present a scalable and efficient attention-based set encoding mechanism that is amenable to mini-batch processing of sets, and capable of updating set representations as data arrives. The proposed method adheres to the required symmetries of invariance and equivariance as well as maintaining MBC for any partition of the input set. We perform extensive experiments and show that our method is computationally efficient and results in rich set encoding representations for set-structured data.

The machine learning (ML) toolbox for estimation of heterogeneous treatment effects from observational data is expanding rapidly, yet many of its algorithms have been evaluated only on a very limited set of semi-synthetic benchmark datasets. In this paper, we investigate current benchmarking practices for ML-based conditional average treatment effect (CATE) estimators, with special focus on empirical evaluation based on the popular semi-synthetic IHDP benchmark. We identify problems with current practice and highlight that semi-synthetic benchmark datasets, which (unlike real-world benchmarks used elsewhere in ML) do not necessarily reflect properties of real data, can systematically favor some algorithms over others -- a fact that is rarely acknowledged but of immense relevance for interpretation of empirical results. Further, we argue that current evaluation metrics evaluate performance only for a small subset of possible use cases of CATE estimators, and discuss alternative metrics relevant for applications in personalized medicine. Additionally, we discuss alternatives for current benchmark datasets, and implications of our findings for benchmarking in CATE estimation.

Federated learning (FL) is a challenging setting for optimization due to the heterogeneity of the data across different clients which gives rise to the client drift phenomenon. In fact, obtaining an algorithm for FL which is uniformly better than simple centralized training has been a major open problem thus far. In this work, we propose a general algorithmic framework, Mime, which i) mitigates client drift and ii) adapts arbitrary centralized optimization algorithms such as momentum and Adam to the cross-device federated learning setting. Mime uses a combination of control-variates and server-level statistics (e.g. momentum) at every client-update step to ensure that each local update mimics that of the centralized method run on iid data. We prove a reduction result showing that Mime can translate the convergence of a generic algorithm in the centralized setting into convergence in the federated setting. Further, we show that when combined with momentum based variance reduction, Mime is provably faster than any centralized method--the first such result. We also perform a thorough experimental exploration of Mime's performance on real world datasets.

Maximum cardinality bipartite matching is an important graph optimization problem with several applications. For instance, maximum cardinality matching in a $\delta$-disc graph can be used in the computation of the bottleneck matching as well as the $\infty$-Wasserstein and the Lévy-Prokhorov distances between probability distributions. For any point sets $A, B \subset \mathbb{R}^2$, the $\delta$-disc graph is a bipartite graph formed by connecting every pair of points $(a,b) \in A\times B$ by an edge if the Euclidean distance between them is at most $\delta$. Using the classical Hopcroft-Karp algorithm, a maximum-cardinality matching on any $\delta$-disc graph can be found in $\tilde{O}(n^{3/2})$ time.~\footnote{We use $\tilde{O}(\cdot)$ to suppress poly-logarithmic terms in the complexity.} In this paper, we present a simplification of a recent algorithm (Lahn and Raghvendra, JoCG 2021) for the maximum cardinality matching problem and describe how a maximum cardinality matching in a $\delta$-disc graph can be computed asymptotically faster than $O(n^{3/2})$ time for any moderately dense point set. As applications, we show that if $A$ and $B$ are point sets drawn uniformly at random from a unit square, an exact bottleneck matching can be computed in $\tilde{O}(n^{4/3})$ time. On the other hand, experiments suggest that the Hopcroft-Karp algorithm seems to take roughly $\Theta (n^{3/2})$ time for this case. This translates to substantial improvements in execution time for larger inputs.

Inspired by human intelligence, the research on online continual learning aims to push the limits of the machine learning models to constantly learn from sequentially encountered tasks, with the data from each task being observed in an online fashion. Though recent studies have achieved remarkable progress in improving the online continual learning performance empowered by the deep neural networks-based models, many of today's approaches still suffer a lot from catastrophic forgetting, a persistent challenge for continual learning. In this paper, we present a novel method which attempts to mitigate catastrophic forgetting in online continual learning from a new perspective, i.e., neuron calibration. In particular, we model the neurons in the deep neural networks-based models as calibrated units under a general formulation. Then we formalize a learning framework to effectively train the calibrated model, where neuron calibration could give ubiquitous benefit to balance the stability and plasticity of online continual learning algorithms through influencing both their forward inference path and backward optimization path. Our proposed formulation for neuron calibration is lightweight and applicable to general feed-forward neural networks-based models. We perform extensive experiments to evaluate our method on four benchmark continual learning datasets. The results show that neuron calibration plays a vital role in improving online continual learning performance and our method could substantially improve the state-of-the-art performance on all~the~evaluated~datasets.

We study the off-policy evaluation (OPE) problem in reinforcement learning with linear function approximation, which aims to estimate the value function of a target policy based on the offline data collected by a behavior policy. We propose to incorporate the variance information of the value function to improve the sample efficiency of OPE. More specifically, for time-inhomogeneous episodic linear Markov decision processes (MDPs), we propose an algorithm, \texttt{VA-OPE}, which uses the estimated variance of the value function to reweight the Bellman residual in Fitted Q-Iteration. We show that our algorithm achieves a tighter error bound than the best-known result. We also provide a fine-grained characterization of the distribution shift between the behavior policy and the target policy. Extensive numerical experiments corroborate our theory.

We study the problem of sparse tensor principal component analysis: given a tensor $\pmb Y = \pmb W + \lambda x^{\otimes p}$ with $\pmb W \in \otimes^p \mathbb{R}^n$ having i.i.d. Gaussian entries, the goal is to recover the $k$-sparse unit vector $x \in \mathbb{R}^n$. The model captures both sparse PCA (in its Wigner form) and tensor PCA. For the highly sparse regime of $k \leq \sqrt{n}$, we present a family of algorithms that smoothly interpolates between a simple polynomial-time algorithm and the exponential-time exhaustive search algorithm. For any $1 \leq t \leq k$, our algorithms recovers the sparse vector for signal-to-noise ratio $\lambda \geq \tilde{\mathcal{O}} (\sqrt{t} \cdot (k/t)^{p/2})$ in time $\tilde{\mathcal{O}}(n^{p+t})$, capturing the state-of-the-art guarantees for the matrix settings (in both the polynomial-time and sub-exponential time regimes). Our results naturally extend to the case of $r$ distinct $k$-sparse signals with disjoint supports, with guarantees that are independent of the number of spikes. Even in the restricted case of sparse PCA, known algorithms only recover the sparse vectors for $\lambda \geq \tilde{\mathcal{O}}(k \cdot r)$ while our algorithms require $\lambda \geq \tilde{\mathcal{O}}(k)$. Finally, by analyzing the low-degree likelihood ratio, we complement these algorithmic results with rigorous evidence illustrating the trade-offs between signal-to-noise ratio and running time. This lower bound captures the known lower bounds for both sparse PCA and tensor PCA. In this general model, we observe a more intricate three-way trade-off between the number of samples $n$, the sparsity $k$, and the tensor power $p$.

In view of training increasingly complex learning architectures, we establish a nonsmooth implicit function theorem with an operational calculus. Our result applies to most practical problems (i.e., definable problems) provided that a nonsmooth form of the classical invertibility condition is fulfilled. This approach allows for formal subdifferentiation: for instance, replacing derivatives by Clarke Jacobians in the usual differentiation formulas is fully justified for a wide class of nonsmooth problems. Moreover this calculus is entirely compatible with algorithmic differentiation (e.g., backpropagation). We provide several applications such as training deep equilibrium networks, training neural nets with conic optimization layers, or hyperparameter-tuning for nonsmooth Lasso-type models. To show the sharpness of our assumptions, we present numerical experiments showcasing the extremely pathological gradient dynamics one can encounter when applying implicit algorithmic differentiation without any hypothesis.

Various non-trivial spaces are becoming popular for embedding structured data such as graphs, texts, or images. Following spherical and hyperbolic spaces, more general product spaces have been proposed. However, searching for the best configuration of a product space is a resource-intensive procedure, which reduces the practical applicability of the idea. We generalize the concept of product space and introduce an overlapping space that does not have the configuration search problem. The main idea is to allow subsets of coordinates to be shared between spaces of different types (Euclidean, hyperbolic, spherical). As a result, we often need fewer coordinates to store the objects. Additionally, we propose an optimization algorithm that automatically learns the optimal configuration. Our experiments confirm that overlapping spaces outperform the competitors in graph embedding tasks with different evaluation metrics. We also perform an empirical analysis in a realistic information retrieval setup, where we compare all spaces by incorporating them into DSSM. In this case, the proposed overlapping space consistently achieves nearly optimal results without any configuration tuning. This allows for reducing training time, which can be essential in large-scale applications.