NeurIPS 2023 论文列表

Existing online learning algorithms for adversarial Markov Decision Processes achieve $\mathcal{O}(\sqrt{T})$ regret after $T$ rounds of interactions even if the loss functions are chosen arbitrarily by an adversary, with the caveat that the transition function has to be fixed. This is because it has been shown that adversarial transition functions make no-regret learning impossible. Despite such impossibility results, in this work, we develop algorithms that can handle both adversarial losses and adversarial transitions, with regret increasing smoothly in the degree of maliciousness of the adversary. More concretely, we first propose an algorithm that enjoys $\widetilde{\mathcal{O}}(\sqrt{T} + C^{P})$ regret where $C^{P}$ measures how adversarial the transition functions are and can be at most $\mathcal{O}(T)$. While this algorithm itself requires knowledge of $C^{P}$, we further develop a black-box reduction approach that removes this requirement. Moreover, we also show that further refinements of the algorithm not only maintains the same regret bound, but also simultaneously adapts to easier environments (where losses are generated in a certain stochastically constrained manner as in [Jin et al. 2021]) and achieves $\widetilde{\mathcal{O}}(U + \sqrt{UC^{L}} + C^{P})$ regret, where $U$ is some standard gap-dependent coefficient and $C^{L}$ is the amount of corruption on losses.

This work is concerned with the estimation problem of linear model when the sample size is extremely large and the data dimension can vary with the sample size. In this setting, the least square estimator based on the full data is not feasible with limited computational resources. Many existing methods for this problem are based on the sketching technique which uses the sketched data to perform least square estimation. We derive fine-grained lower bounds of the conditional mean squared error for sketching methods. For sampling methods, our lower bound provides an attainable optimal convergence rate. Our result implies that when the dimension is large, there is hardly a sampling method can have a faster convergence rate than the uniform sampling method. To achieve a better statistical performance, we propose a new sketching method based on data averaging. The proposed method reduces the original data to a few averaged observations. These averaged observations still satisfy the linear model and are used to estimate the regression coefficients. The asymptotic behavior of the proposed estimation procedure is studied. Our theoretical results show that the proposed method can achieve a faster convergence rate than the optimal convergence rate for sampling methods. Theoretical and numerical results show that the proposed estimator has good statistical performance as well as low computational cost.

The dramatic growth of global e-commerce has led to a surge in demand for efficient and cost-effective order fulfillment which can increase customers' service levels and sellers' competitiveness. However, managing order fulfillment is challenging due to a series of interdependent online sequential decision-making problems. To clear this hurdle, rather than solving the problems separately as attempted in some recent researches, this paper proposes a method based on multi-agent reinforcement learning to integratively solve the series of interconnected problems, encompassing order handling, packing and pickup, storage, order consolidation, and last-mile delivery. In particular, we model the integrated problem as a Markov game, wherein a team of agents learns a joint policy via interacting with a simulated environment. Since no simulated environment supporting the complete order fulfillment problem exists, we devise Order Fulfillment COoperative mUlti-agent Reinforcement learning Scalable Environment (OFCOURSE) in the OpenAI Gym style, which allows reproduction and re-utilization to build customized applications. By constructing the fulfillment system in OFCOURSE, we optimize a joint policy that solves the integrated problem, facilitating sequential order-wise operations across all fulfillment units and minimizing the total cost of fulfilling all orders within the promised time. With OFCOURSE, we also demonstrate that the joint policy learned by multi-agent reinforcement learning outperforms the combination of locally optimal policies. The source code of OFCOURSE is available at: https://github.com/GitYiheng/ofcourse.

Offline reinforcement learning (RL) aims to learn optimal policies from offline datasets, where the parameterization of policies is crucial but often overlooked. Recently, Diffsuion-QL significantly boosts the performance of offline RL by representing a policy with a diffusion model, whose success relies on a parametrized Markov Chain with hundreds of steps for sampling. However, Diffusion-QL suffers from two critical limitations. 1) It is computationally inefficient to forward and backward through the whole Markov chain during training. 2) It is incompatible with maximum likelihood-based RL algorithms (e.g., policy gradient methods) as the likelihood of diffusion models is intractable. Therefore, we propose efficient diffusion policy (EDP) to overcome these two challenges. EDP approximately constructs actions from corrupted ones at training to avoid running the sampling chain. We conduct extensive experiments on the D4RL benchmark. The results show that EDP can reduce the diffusion policy training time from 5 days to 5 hours on gym-locomotion tasks. Moreover, we show that EDP is compatible with various offline RL algorithms (TD3, CRR, and IQL) and achieves new state-of-the-art on D4RL by large margins over previous methods.

Neural network pruning and quantization techniques are almost as old as neural networks themselves. However, to date, only ad-hoc comparisons between the two have been published. In this paper, we set out to answer the question of which is better: neural network quantization or pruning? By answering this question, we hope to inform design decisions made on neural network hardware going forward. We provide an extensive comparison between the two techniques for compressing deep neural networks. First, we give an analytical comparison of expected quantization and pruning error for general data distributions. Then, we provide lower and upper bounds for the per-layer pruning and quantization error in trained networks and compare these to empirical error after optimization. Finally, we provide an extensive experimental comparison for training 8 large-scale models trained on 3 tasks and provide insights into the representations learned during fine-tuning with quantization and pruning in the loop. Our results show that in most cases quantization outperforms pruning. Only in some scenarios with a very high compression ratio, compression might be beneficial from an accuracy standpoint.

Federated reinforcement learning (FedRL) enables agents to collaboratively train a global policy without sharing their individual data. However, high communication overhead remains a critical bottleneck, particularly for natural policy gradient (NPG) methods, which are second-order. To address this issue, we propose the FedNPG-ADMM framework, which leverages the alternating direction method of multipliers (ADMM) to approximate global NPG directions efficiently. We theoretically demonstrate that using ADMM-based gradient updates reduces communication complexity from $\mathcal{O}({d^{2}})$ to $\mathcal{O}({d})$ at each iteration, where $d$ is the number of model parameters. Furthermore, we show that achieving an $\epsilon$-error stationary convergence requires $\mathcal{O}(\frac{1}{(1-\gamma)^{2}{\epsilon}})$ iterations for discount factor $\gamma$, demonstrating that FedNPG-ADMM maintains the same convergence rate as standard FedNPG. Through evaluation of the proposed algorithms in MuJoCo environments, we demonstrate that FedNPG-ADMM maintains the reward performance of standard FedNPG, and that its convergence rate improves when the number of federated agents increases.

Geometric representation learning of molecules is challenging yet essential for applications in multiple domains. Despite the impressive breakthroughs made by geometric deep learning in various molecular representation learning tasks, effectively capturing complicated geometric features across spatial dimensions is still underexplored due to the significant difficulties in modeling efficient geometric representations and learning the inherent correlation in 3D structural modeling. These include computational inefficiency, underutilization of vectorial embeddings, and limited generalizability to integrate various geometric properties. To address the raised concerns, we introduce an efficient and effective framework, Scalable Vector Network (SaVeNet), designed to accommodate a range of geometric requirements without depending on costly embeddings. In addition, the proposed framework scales effectively with introduced direction noise. Theoretically, we analyze the desired properties (i.e., invariance and equivariant) and framework efficiency of the SaVeNet. Empirically, we conduct a comprehensive series of experiments to evaluate the efficiency and expressiveness of the proposed model. Our efficiency-focused experiments underscore the model's empirical superiority over existing methods. Experimental results on synthetic and real-world datasets demonstrate the expressiveness of our model, which achieves state-of-the-art performance across various tasks within molecular representation learning.

Recently, diffusion models have achieved great success in generative tasks. Sampling from diffusion models is equivalent to solving the reverse diffusion stochastic differential equations (SDEs) or the corresponding probability flow ordinary differential equations (ODEs). In comparison, SDE-based solvers can generate samples of higher quality and are suited for image translation tasks like stroke-based synthesis. During inference, however, existing SDE-based solvers are severely constrained by the efficiency-effectiveness dilemma. Our investigation suggests that this is because the Gaussian assumption in the reverse transition kernel is frequently violated (even in the case of simple mixture data) given a limited number of discretization steps. To overcome this limitation, we introduce a novel class of SDE-based solvers called \emph{Gaussian Mixture Solvers (GMS)} for diffusion models. Our solver estimates the first three-order moments and optimizes the parameters of a Gaussian mixture transition kernel using generalized methods of moments in each step during sampling. Empirically, our solver outperforms numerous SDE-based solvers in terms of sample quality in image generation and stroke-based synthesis in various diffusion models, which validates the motivation and effectiveness of GMS. Our code is available at https://github.com/Guohanzhong/GMS.

In diffusion models, UNet is the most popular network backbone, since its long skip connects (LSCs) to connect distant network blocks can aggregate long-distant information and alleviate vanishing gradient. Unfortunately, UNet often suffers from unstable training in diffusion models which can be alleviated by scaling its LSC coefficients smaller. However, theoretical understandings of the instability of UNet in diffusion models and also the performance improvement of LSC scaling remain absent yet. To solve this issue, we theoretically show that the coefficients of LSCs in UNet have big effects on the stableness of the forward and backward propagation and robustness of UNet. Specifically, the hidden feature and gradient of UNet at any layer can oscillate and their oscillation ranges are actually large which explains the instability of UNet training. Moreover, UNet is also provably sensitive to perturbed input, and predicts an output distant from the desired output, yielding oscillatory loss and thus oscillatory gradient. Besides, we also observe the theoretical benefits of the LSC coefficient scaling of UNet in the stableness of hidden features and gradient and also robustness. Finally, inspired by our theory, we propose an effective coefficient scaling framework ScaleLong that scales the coefficients of LSC in UNet and better improve the training stability of UNet. Experimental results on CIFAR10, CelebA, ImageNet and COCO show that our methods are superior to stabilize training, and yield about 1.5x training acceleration on different diffusion models with UNet or UViT backbones.

NetHack is known as the frontier of reinforcement learning research where learning-based methods still need to catch up to rule-based solutions. One of the promising directions for a breakthrough is using pre-collected datasets similar to recent developments in robotics, recommender systems, and more under the umbrella of offline reinforcement learning (ORL). Recently, a large-scale NetHack dataset was released; while it was a necessary step forward, it has yet to gain wide adoption in the ORL community. In this work, we argue that there are three major obstacles for adoption: tool-wise, implementation-wise, and benchmark-wise. To address them, we develop an open-source library that provides workflow fundamentals familiar to the ORL community: pre-defined D4RL-style tasks, uncluttered baseline implementations, and reliable evaluation tools with accompanying configs and logs synced to the cloud.

In recent years, prompt tuning has sparked a research surge in adapting pre-trained models. Unlike the unified pre-training strategy employed in the language field, the graph field exhibits diverse pre-training strategies, posing challenges in designing appropriate prompt-based tuning methods for graph neural networks. While some pioneering work has devised specialized prompting functions for models that employ edge prediction as their pre-training tasks, these methods are limited to specific pre-trained GNN models and lack broader applicability. In this paper, we introduce a universal prompt-based tuning method called Graph Prompt Feature (GPF) for pre-trained GNN models under any pre-training strategy. GPF operates on the input graph's feature space and can theoretically achieve an equivalent effect to any form of prompting function. Consequently, we no longer need to illustrate the prompting function corresponding to each pre-training strategy explicitly. Instead, we employ GPF to obtain the prompted graph for the downstream task in an adaptive manner. We provide rigorous derivations to demonstrate the universality of GPF and make guarantee of its effectiveness. The experimental results under various pre-training strategies indicate that our method performs better than fine-tuning, with an average improvement of about 1.4% in full-shot scenarios and about 3.2% in few-shot scenarios. Moreover, our method significantly outperforms existing specialized prompt-based tuning methods when applied to models utilizing the pre-training strategy they specialize in. These numerous advantages position our method as a compelling alternative to fine-tuning for downstream adaptations.

Maximizing submodular functions has been increasingly used in many applications of machine learning, such as data summarization, recommendation systems, and feature selection. Moreover, there has been a growing interest in both submodular maximization and dynamic algorithms. In 2020, Monemizadeh and Lattanzi, Mitrovic, Norouzi-Fard, Tarnawski, and Zadimoghaddam initiated developing dynamic algorithms for the monotone submodular maximization problem under the cardinality constraint $k$. In 2022, Chen and Peng studied the complexity of this problem and raised an important open question: "\emph{Can we extend [fully dynamic] results (algorithm or hardness) to non-monotone submodular maximization?}". We affirmatively answer their question by demonstrating a reduction from maximizing a non-monotone submodular function under the cardinality constraint $k$ to maximizing a monotone submodular function under the same constraint. Through this reduction, we obtain the first dynamic algorithms to solve the non-monotone submodular maximization problem under the cardinality constraint $k$. Our algorithms maintain an $(8+\epsilon)$-approximate of the solution and use expected amortized $O(\epsilon^{-3}k^3\log^3(n)\log(k))$ or $O(\epsilon^{-1}k^2\log^3(k))$ oracle queries per update, respectively. Furthermore, we showcase the benefits of our dynamic algorithm for video summarization and max-cut problems on several real-world data sets.

Multi-agent reinforcement learning (MARL) has primarily focused on solving a single task in isolation, while in practice the environment is often evolving, leaving many related tasks to be solved. In this paper, we investigate the benefits of meta-learning in solving multiple MARL tasks collectively. We establish the first line of theoretical results for meta-learning in a wide range of fundamental MARL settings, including learning Nash equilibria in two-player zero-sum Markov games and Markov potential games, as well as learning coarse correlated equilibria in general-sum Markov games. Under natural notions of task similarity, we show that meta-learning achieves provable sharper convergence to various game-theoretical solution concepts than learning each task separately. As an important intermediate step, we develop multiple MARL algorithms with initialization-dependent convergence guarantees. Such algorithms integrate optimistic policy mirror descents with stage-based value updates, and their refined convergence guarantees (nearly) recover the best known results even when a good initialization is unknown. To our best knowledge, such results are also new and might be of independent interest. We further provide numerical simulations to corroborate our theoretical findings.

We present RoboHive, a comprehensive software platform and ecosystem for research in the field of Robot Learning and Embodied Artificial Intelligence. Our platform encompasses a diverse range of pre-existing and novel environments, including dexterous manipulation with the Shadow Hand, whole-arm manipulation tasks with Franka and Fetch robots, quadruped locomotion, among others. Included environments are organized within and cover multiple domains such as hand manipulation, locomotion, multi-task, multi-agent, muscles, etc. In comparison to prior works, RoboHive offers a streamlined and unified task interface taking dependency on only a minimal set of well-maintained packages, features tasks with high physics fidelity and rich visual diversity, and supports common hardware drivers for real-world deployment. The unified interface of RoboHive offers a convenient and accessible abstraction for algorithmic research in imitation, reinforcement, multi-task, and hierarchical learning. Furthermore, RoboHive includes expert demonstrations and baseline results for most environments, providing a standard for benchmarking and comparisons. Details: https://sites.google.com/view/robohive

Reward shaping is an effective technique for integrating domain knowledge into reinforcement learning (RL). However, traditional approaches like potential-based reward shaping totally rely on manually designing shaping reward functions, which significantly restricts exploration efficiency and introduces human cognitive biases. While a number of RL methods have been proposed to boost exploration by designing an intrinsic reward signal as exploration bonus. Nevertheless, these methods heavily rely on the count-based episodic term in their exploration bonus which falls short in scalability. To address these limitations, we propose a general end-to-end potential-based exploration bonus for deep RL via potentials of state discrepancy, which motivates the agent to discover novel states and provides them with denser rewards without manual intervention. Specifically, we measure the novelty of adjacent states by calculating their distance using the bisimulation metric-based potential function, which enhances agent's exploration and ensures policy invariance. In addition, we offer a theoretical guarantee on our inverse dynamic bisimulation metric, bounding the value difference and ensuring that the agent explores states with higher TD error, thus significantly improving training efficiency. The proposed approach is named \textbf{LIBERTY} (exp\textbf{L}oration v\textbf{I}a \textbf{B}isimulation m\textbf{E}t\textbf{R}ic-based s\textbf{T}ate discrepanc\textbf{Y}) which is comprehensively evaluated on the MuJoCo and the Arcade Learning Environments. Extensive experiments have verified the superiority and scalability of our algorithm compared with other competitive methods.

Diffusion models have gained significant attention in the realm of image generation due to their exceptional performance. Their success has been recently expanded to text generation via generating all tokens within a sequence concurrently. However, natural language exhibits a far more pronounced sequential dependency in comparison to images, and the majority of existing language models are trained with a left-to-right auto-regressive approach. To account for the inherent sequential characteristic of natural language, we introduce Auto-Regressive Diffusion (AR-Diffusion). AR-Diffusion ensures that the generation of tokens on the right depends on the generated ones on the left, a mechanism achieved through employing a dynamic number of denoising steps that vary based on token position. This results in tokens on the left undergoing fewer denoising steps than those on the right, thereby enabling them to generate earlier and subsequently influence the generation of tokens on the right. In a series of experiments on various text generation tasks, including text summarization, machine translation, and common sense generation, AR-Diffusion clearly demonstrated its superiority over existing diffusion language models and that it can be $100\times\sim600\times$ faster when achieving comparable results. Our code is available at https://github.com/microsoft/ProphetNet/tree/master/AR-diffusion.

Uncertainty estimation aims to evaluate the confidence of a trained deep neural network. However, existing uncertainty estimation approaches rely on low-dimensional distributional assumptions and thus suffer from the high dimensionality of latent features. Existing approaches tend to focus on uncertainty on discrete classification probabilities, which leads to poor generalizability to uncertainty estimation for other tasks. Moreover, most of the literature requires seeing the out-of-distribution (OOD) data in the training for better estimation of uncertainty, which limits the uncertainty estimation performance in practice because the OOD data are typically unseen. To overcome these limitations, we propose a new framework using data-adaptive high-dimensional hypothesis testing for uncertainty estimation, which leverages the statistical properties of the feature representations. Our method directly operates on latent representations and thus does not require retraining the feature encoder under a modified objective. The test statistic relaxes the feature distribution assumptions to high dimensionality, and it is more discriminative to uncertainties in the latent representations. We demonstrate that encoding features with Bayesian neural networks can enhance testing performance and lead to more accurate uncertainty estimation. We further introduce a family-wise testing procedure to determine the optimal threshold of OOD detection, which minimizes the false discovery rate (FDR). Extensive experiments validate the satisfactory performance of our framework on uncertainty estimation and task-specific prediction over a variety of competitors. The experiments on the OOD detection task also show satisfactory performance of our method when the OOD data are unseen in the training. Codes are available at https://github.com/HKU-MedAI/bnn_uncertainty.

In partial label learning (PLL), each sample is associated with a group of candidate labels, among which only one label is correct. The key of PLL is to disambiguate the candidate label set to find the ground-truth label. To this end, we first construct a constrained regression model to capture the confidence of the candidate labels, and multiply the label confidence matrix by its transpose to build a second-order similarity matrix, whose elements indicate the pairwise similarity relationships of samples globally. Then we develop a semantic dissimilarity matrix by considering the complement of the intersection of the candidate label set, and further propagate the initial dissimilarity relationships to the whole data set by leveraging the local geometric structure of samples. The similarity and dissimilarity matrices form an adversarial relationship, which is further utilized to shrink the solution space of the label confidence matrix and promote the dissimilarity matrix. We finally extend the proposed model to a kernel version to exploit the non-linear structure of samples and solve the proposed model by the inexact augmented Lagrange multiplier method. By exploiting the adversarial prior, the proposed method can significantly outperform state-of-the-art PLL algorithms when evaluated on 10 artificial and 7 real-world partial label data sets. We also prove the effectiveness of our method with some theoretical guarantees. The code is publicly available at https://github.com/Yangfc-ML/DPCLS.

Understanding which concepts models can and cannot represent has been fundamental to many tasks: from effective and responsible use of models to detecting out of distribution data. We introduce Gaussian process probes (GPP), a unified and simple framework for probing and measuring uncertainty about concepts represented by models. As a Bayesian extension of linear probing methods, GPP asks what kind of distribution over classifiers (of concepts) is induced by the model. This distribution can be used to measure both what the model represents and how confident the probe is about what the model represents. GPP can be applied to any pre-trained model with vector representations of inputs (e.g., activations). It does not require access to training data, gradients, or the architecture. We validate GPP on datasets containing both synthetic and real images. Our experiments show it can (1) probe a model's representations of concepts even with a very small number of examples, (2) accurately measure both epistemic uncertainty (how confident the probe is) and aleatory uncertainty (how fuzzy the concepts are to the model), and (3) detect out of distribution data using those uncertainty measures as well as classic methods do. By using Gaussian processes to expand what probing can offer, GPP provides a data-efficient, versatile and uncertainty-aware tool for understanding and evaluating the capabilities of machine learning models.