ICLR 2022 Papers

In this paper, we establish the first finite-time convergence result of the actor-critic algorithm for fully decentralized multi-agent reinforcement learning (MARL) problems with average reward. In this problem, a set of $N$ agents work cooperatively to maximize the global average reward through interacting with their neighbors over a communication network. We consider a practical MARL setting, where the rewards and actions of each agent are only known to itself, and the knowledge of joint actions of the agents is not assumed. Toward this end, we propose a mini-batch Markovian sampled fully decentralized actor-critic algorithm and analyze its finite-time convergence and sample complexity. We show that the sample complexity of this algorithm is $\mathcal{O}(N^{2}/\epsilon^{2}\log(N/\epsilon))$. Interestingly, this sample complexity bound matches that of the state-of-the-art single-agent actor-critic algorithms for reinforcement learning.

Generating the periodic structure of stable materials is a long-standing challenge for the material design community. This task is difficult because stable materials only exist in a low-dimensional subspace of all possible periodic arrangements of atoms: 1) the coordinates must lie in the local energy minimum defined by quantum mechanics, and 2) global stability also requires the structure to follow the complex, yet specific bonding preferences between different atom types. Existing methods fail to incorporate these factors and often lack proper invariances. We propose a Crystal Diffusion Variational Autoencoder (CDVAE) that captures the physical inductive bias of material stability. By learning from the data distribution of stable materials, the decoder generates materials in a diffusion process that moves atomic coordinates towards a lower energy state and updates atom types to satisfy bonding preferences between neighbors. Our model also explicitly encodes interactions across periodic boundaries and respects permutation, translation, rotation, and periodic invariances. We significantly outperform past methods in three tasks: 1) reconstructing the input structure, 2) generating valid, diverse, and realistic materials, and 3) generating materials that optimize a specific property. We also provide several standard datasets and evaluation metrics for the broader machine learning community.

Many gradient-based meta-learning methods assume a set of parameters that do not participate in inner-optimization, which can be considered as hyperparameters. Although such hyperparameters can be optimized using the existing gradient-based hyperparameter optimization (HO) methods, they suffer from the following issues. Unrolled differentiation methods do not scale well to high-dimensional hyperparameters or horizon length, Implicit Function Theorem (IFT) based methods are restrictive for online optimization, and short horizon approximations suffer from short horizon bias. In this work, we propose a novel HO method that can overcome these limitations, by approximating the second-order term with knowledge distillation. Specifically, we parameterize a single Jacobian-vector product (JVP) for each HO step and minimize the distance from the true second-order term. Our method allows online optimization and also is scalable to the hyperparameter dimension and the horizon length. We demonstrate the effectiveness of our method on three different meta-learning methods and two benchmark datasets.

In vitro cellular experimentation with genetic interventions, using for example CRISPR technologies, is an essential step in early-stage drug discovery and target validation that serves to assess initial hypotheses about causal associations between biological mechanisms and disease pathologies. With billions of potential hypotheses to test, the experimental design space for in vitro genetic experiments is extremely vast, and the available experimental capacity - even at the largest research institutions in the world - pales in relation to the size of this biological hypothesis space. Machine learning methods, such as active and reinforcement learning, could aid in optimally exploring the vast biological space by integrating prior knowledge from various information sources as well as extrapolating to yet unexplored areas of the experimental design space based on available data. However, there exist no standardised benchmarks and data sets for this challenging task and little research has been conducted in this area to date. Here, we introduce GeneDisco, a benchmark suite for evaluating active learning algorithms for experimental design in drug discovery. GeneDisco contains a curated set of multiple publicly available experimental data sets as well as open-source implementations of state-of-the-art active learning policies for experimental design and exploration.

3D shape analysis has been widely explored in the era of deep learning. Numerous models have been developed for various 3D data representation formats, e.g., MeshCNN for meshes, PointNet for point clouds and VoxNet for voxels. In this study, we present Representation-Agnostic Shape Fields (RASF), a generalizable and computation-efficient shape embedding module for 3D deep learning. RASF is implemented with a learnable 3D grid with multiple channels to store local geometry. Based on RASF, shape embeddings for various 3D shape representations (point clouds, meshes and voxels) are retrieved by coordinate indexing. While there are multiple ways to optimize the learnable parameters of RASF, we provide two effective schemes among all in this paper for RASF pre-training: shape reconstruction and normal estimation. Once trained, RASF becomes a plug-and-play performance booster with negligible cost. Extensive experiments on diverse 3D representation formats, networks and applications, validate the universal effectiveness of the proposed RASF. Code and pre-trained models are publicly available\footnote{\url{https://github.com/seanywang0408/RASF}}.

We perform systematically and fairly controlled experiments with the 6-layer Transformer to investigate the hardness in conditional-language-modeling languages which have been traditionally considered morphologically rich (AR and RU) and poor (ZH). We evaluate through statistical comparisons across 30 possible language directions from the 6 languages of the United Nations Parallel Corpus across 5 data sizes on 3 representation levels --- character, byte, and word. Results show that performance is relative to the representation granularity of each of the languages, not to the language as a whole. On the character and byte levels, we are able to eliminate statistically significant performance disparity, hence demonstrating that a language cannot be intrinsically hard. The disparity that mirrors the morphological complexity hierarchy is shown to be a byproduct of word segmentation. Evidence from data statistics, along with the fact that word segmentation is qualitatively indeterminate, renders a decades-long debate on morphological complexity (unless it is being intentionally modeled in a word-based, meaning-driven context) irrelevant in the context of computing. The intent of our work is to help effect more objectivity and adequacy in evaluation as well as fairness and inclusivity in experimental setup in the area of language and computing so to uphold diversity in Machine Learning and Artificial Intelligence research. Multilinguality is real and relevant in computing not due to canonical, structural linguistic concepts such as morphology or "words" in our minds, but rather standards related to internationalization and localization, such as character encoding --- something which has thus far been sorely overlooked in our discourse and curricula.

Due to spurious correlations, machine learning systems often fail to generalize to environments whose distributions differ from the ones used at training time. Prior work addressing this, either explicitly or implicitly, attempted to find a data representation that has an invariant relationship with the target. This is done by leveraging a diverse set of training environments to reduce the effect of spurious features and build an invariant predictor. However, these methods have generalization guarantees only when both data representation and classifiers come from a linear model class. We propose invariant Causal Representation Learning (iCaRL), an approach that enables out-of-distribution (OOD) generalization in the nonlinear setting (i.e., nonlinear representations and nonlinear classifiers). It builds upon a practical and general assumption: the prior over the data representation (i.e., a set of latent variables encoding the data) given the target and the environment belongs to general exponential family distributions, i.e., a more flexible conditionally non-factorized prior that can actually capture complicated dependences between the latent variables. Based on this, we show that it is possible to identify the data representation up to simple transformations. We also show that all direct causes of the target can be fully discovered, which further enables us to obtain generalization guarantees in the nonlinear setting. Experiments on both synthetic and real-world datasets demonstrate that our approach outperforms a variety of baseline methods.

The performance of existing text style transfer models is severely limited by the non-parallel datasets on which the models are trained. In non-parallel datasets, no direct mapping exists between sentences of the source and target style; the style transfer models thus only receive weak supervision of the target sentences during training, which often leads the model to discard too much style-independent information, or utterly fail to transfer the style. In this work, we propose LaMer, a novel text style transfer framework based on large-scale language models. LaMer first mines the roughly parallel expressions in the non-parallel datasets with scene graphs, and then employs MLE training, followed by imitation learning refinement, to leverage the intrinsic parallelism within the data. On two benchmark tasks (sentiment & formality transfer) and a newly proposed challenging task (political stance transfer), our model achieves qualitative advances in transfer accuracy, content preservation, and fluency. Further empirical and human evaluations demonstrate that our model not only makes training more efficient, but also generates more readable and diverse expressions than previous models.

Emphatic temporal difference (ETD) learning (Sutton et al., 2016) is a successful method to conduct the off-policy value function evaluation with function approximation. Although ETD has been shown to converge asymptotically to a desirable value function, it is well-known that ETD often encounters a large variance so that its sample complexity can increase exponentially fast with the number of iterations. In this work, we propose a new ETD method, called PER-ETD (i.e., PEriodically Restarted-ETD), which restarts and updates the follow-on trace only for a finite period for each iteration of the evaluation parameter. Further, PER-ETD features a design of the logarithmical increase of the restart period with the number of iterations, which guarantees the best trade-off between the variance and bias and keeps both vanishing sublinearly. We show that PER-ETD converges to the same desirable fixed point as ETD, but improves the exponential sample complexity of ETD to be polynomials. Our experiments validate the superior performance of PER-ETD and its advantage over ETD.

This work addresses one-shot set and graph generation, and, more specifically, the parametrization of probabilistic decoders that map a vector-shaped prior to a distribution over sets or graphs. Sets and graphs are most commonly generated by first sampling points i.i.d. from a normal distribution, and then processing these points along with the prior vector using Transformer layers or Graph Neural Networks. This architecture is designed to generate exchangeable distributions, i.e., all permutations of the generated outputs are equally likely. We however show that it only optimizes a proxy to the evidence lower bound, which makes it hard to train. We then study equivariance in generative settings and show that non-exchangeable methods can still achieve permutation equivariance. Using this result, we introduce Top-n creation, a differentiable generation mechanism that uses the latent vector to select the most relevant points from a trainable reference set. Top-n can replace i.i.d. generation in any Variational Autoencoder or Generative Adversarial Network. Experimentally, our method outperforms i.i.d. generation by 15% at SetMNIST reconstruction, by 33% at object detection on CLEVR, generates sets that are 74% closer to the true distribution on a synthetic molecule-like dataset, and generates more valid molecules on QM9.

In recent years, implicit deep learning has emerged as a method to increase the depth of deep neural networks. While their training is memory-efficient, they are still significantly slower to train than their explicit counterparts. In Deep Equilibrium Models~(DEQs), the training is performed as a bi-level problem, and its computational complexity is partially driven by the iterative inversion of a huge Jacobian matrix. In this paper, we propose a novel strategy to tackle this computational bottleneck from which many bi-level problems suffer. The main idea is to use the quasi-Newton matrices from the forward pass to efficiently approximate the inverse Jacobian matrix in the direction needed for the gradient computation. We provide a theorem that motivates using our method with the original forward algorithms. In addition, by modifying these forward algorithms, we further provide theoretical guarantees that our method asymptotically estimates the true implicit gradient. We empirically study this approach in many settings, ranging from hyperparameter optimization to large Multiscale DEQs applied to CIFAR and ImageNet. We show that it reduces the computational cost of the backward pass by up to two orders of magnitude. All this is achieved while retaining the excellent performance of the original models in hyperparameter optimization and on CIFAR, and giving encouraging and competitive results on ImageNet.

Computing the matrix square root or its inverse in a differentiable manner is important in a variety of computer vision tasks. Previous methods either adopt the Singular Value Decomposition (SVD) to explicitly factorize the matrix or use the Newton-Schulz iteration (NS iteration) to derive the approximate solution. However, both methods are not computationally efficient enough in either the forward pass or in the backward pass. In this paper, we propose two more efficient variants to compute the differentiable matrix square root. For the forward propagation, one method is to use Matrix Taylor Polynomial (MTP), and the other method is to use Matrix Pad\'e Approximants (MPA). The backward gradient is computed by iteratively solving the continuous-time Lyapunov equation using the matrix sign function. Both methods yield considerable speed-up compared with the SVD or the Newton-Schulz iteration. Experimental results on the de-correlated batch normalization and second-order vision transformer demonstrate that our methods can also achieve competitive and even slightly better performances. The code is available at \href{https://github.com/KingJamesSong/FastDifferentiableMatSqrt}{https://github.com/KingJamesSong/FastDifferentiableMatSqrt}.

Neuromorphic neural network processors, in the form of compute-in-memory crossbar arrays of memristors, or in the form of subthreshold analog and mixed-signal ASICs, promise enormous advantages in compute density and energy efficiency for NN-based ML tasks. However, these technologies are prone to computational non-idealities, due to process variation and intrinsic device physics. This degrades the task performance of networks deployed to the processor, by introducing parameter noise into the deployed model. While it is possible to calibrate each device, or train networks individually for each processor, these approaches are expensive and impractical for commercial deployment. Alternative methods are therefore needed to train networks that are inherently robust against parameter variation, as a consequence of network architecture and parameters. We present a new network training algorithm that attacks network parameters during training, and promotes robust performance during inference in the face of random parameter variation. Our approach introduces a loss regularization term that penalizes the susceptibility of a network to weight perturbation. We compare against previous approaches for producing parameter insensitivity such as dropout, weight smoothing and introducing parameter noise during training. We show that our approach produces models that are more robust to random mismatch-induced parameter variation as well as to targeted parameter variation. Our approach finds minima in flatter locations in the weight-loss landscape compared with other approaches, highlighting that the networks found by our technique are less sensitive to parameter perturbation. Our work provides an approach to deploy neural network architectures to inference devices that suffer from computational non-idealities, with minimal loss of performance. This method will enable deployment at scale to novel energy-efficient computational substrates, promoting cheaper and more prevalent edge inference.

For many differentially private algorithms, such as the prominent noisy stochastic gradient descent (DP-SGD), the analysis needed to bound the privacy leakage of a single training run is well understood. However, few studies have reasoned about the privacy leakage resulting from the multiple training runs needed to fine tune the value of the training algorithm’s hyperparameters. In this work, we first illustrate how simply setting hyperparameters based on non-private training runs can leak private information. Motivated by this observation, we then provide privacy guarantees for hyperparameter search procedures within the framework of Renyi Differential Privacy. Our results improve and extend the work of Liu and Talwar (STOC 2019). Our analysis supports our previous observation that tuning hyperparameters does indeed leak private information, but we prove that, under certain assumptions, this leakage is modest, as long as each candidate training run needed to select hyperparameters is itself differentially private.