NeurIPS 2023 论文列表

This paper focuses on the data-insufficiency problem in multi-task learning within an episodic training setup. Specifically, we explore the potential of heterogeneous information across tasks and meta-knowledge among episodes to effectively tackle each task with limited data. Existing meta-learning methods often fail to take advantage of crucial heterogeneous information in a single episode, while multi-task learning models neglect reusing experience from earlier episodes. To address the problem of insufficient data, we develop Heterogeneous Neural Processes (HNPs) for the episodic multi-task setup. Within the framework of hierarchical Bayes, HNPs effectively capitalize on prior experiences as meta-knowledge and capture task-relatedness among heterogeneous tasks, mitigating data-insufficiency. Meanwhile, transformer-structured inference modules are designed to enable efficient inferences toward meta-knowledge and task-relatedness. In this way, HNPs can learn more powerful functional priors for adapting to novel heterogeneous tasks in each meta-test episode. Experimental results show the superior performance of the proposed HNPs over typical baselines, and ablation studies verify the effectiveness of the designed inference modules.

This study considers a novel problem setting, referred to as \textit{bandit task assignment}, that incorporates the processing time of each task in the bandit setting. In this problem setting, a player sequentially chooses a set of tasks to start so that the set of processing tasks satisfies a given combinatorial constraint. The reward and processing time for each task follow unknown distributions, values of which are revealed only after the task has been completed. The problem generalizes the stochastic combinatorial semi-bandit problem and the budget-constrained bandit problem. For this problem setting, we propose an algorithm based on upper confidence bounds~(UCB) combined with a phased-update approach. The proposed algorithm admits a gap-dependent regret upper bound of $O(MN(1/\Delta){\log T})$ and a gap-free regret upper bound of $\tilde{O}( \sqrt{MNT} )$, where $N$ is the number of the tasks, $M$ is the maximum number of tasks run at the same time, $T$ is the time horizon, and $\Delta$ is the gap between expected per-round rewards of the optimal and best suboptimal sets of tasks. These regret bounds nearly match lower bounds.

Vision-language pretrained models have seen remarkable success, but their application to safety-critical settings is limited by their lack of interpretability. To improve the interpretability of vision-language models such as CLIP, we propose a multi-modal information bottleneck (M2IB) approach that learns latent representations that compress irrelevant information while preserving relevant visual and textual features. We demonstrate how M2IB can be applied to attribution analysis of vision-language pretrained models, increasing attribution accuracy and improving the interpretability of such models when applied to safety-critical domains such as healthcare. Crucially, unlike commonly used unimodal attribution methods, M2IB does not require ground truth labels, making it possible to audit representations of vision-language pretrained models when multiple modalities but no ground-truth data is available. Using CLIP as an example, we demonstrate the effectiveness of M2IB attribution and show that it outperforms gradient-based, perturbation-based, and attention-based attribution methods both qualitatively and quantitatively.

Neural MMO 2.0 is a massively multi-agent and multi-task environment for reinforcement learning research. This version features a novel task-system that broadens the range of training settings and poses a new challenge in generalization: evaluation on and against tasks, maps, and opponents never seen during training. Maps are procedurally generated with 128 agents in the standard setting and 1-1024 supported overall. Version 2.0 is a complete rewrite of its predecessor with three-fold improved performance, effectively addressing simulation bottlenecks in online training. Enhancements to compatibility enable training with standard reinforcement learning frameworks designed for much simpler environments. Neural MMO 2.0 is free and open-source with comprehensive documentation available at neuralmmo.github.io and an active community Discord. To spark initial research on this new platform, we are concurrently running a competition at NeurIPS 2023.

Prior research on policy adaptation in multi-agent games has often relied on online interaction with the target agent in training, which can be expensive and impractical in real-world scenarios. Inspired by recent progress in offline reinforcement learn- ing, this paper studies offline policy adaptation, which aims to utilize the target agent’s behavior data to exploit its weakness or enable effective cooperation. We investigate its distinct challenges of distributional shift and risk-free deviation, and propose a novel learning objective, conservative offline adaptation, that optimizes the worst-case performance against any dataset consistent proxy models. We pro- pose an efficient algorithm called Constrained Self-Play (CSP) that incorporates dataset information into regularized policy learning. We prove that CSP learns a near-optimal risk-free offline adaptation policy upon convergence. Empirical results demonstrate that CSP outperforms non-conservative baselines in various environments, including Maze, predator-prey, MuJoCo, and Google Football.

Neural Radiance Fields (NeRF) have become an increasingly popular representation to capture high-quality appearance and shape of scenes and objects. However, learning generalizable NeRF priors over categories of scenes or objects has been challenging due to the high dimensionality of network weight space. To address the limitations of existing work on generalization, multi-view consistency and to improve quality, we propose HyP-NeRF, a latent conditioning method for learning generalizable category-level NeRF priors using hypernetworks. Rather than using hypernetworks to estimate only the weights of a NeRF, we estimate both the weights and the multi-resolution hash encodings resulting in significant quality gains. To improve quality even further, we incorporate a denoise and finetune strategy that denoises images rendered from NeRFs estimated by the hypernetwork and finetunes it while retaining multiview consistency. These improvements enable us to use HyP-NeRF as a generalizable prior for multiple downstream tasks including NeRF reconstruction from single-view or cluttered scenes and text-to-NeRF. We provide qualitative comparisons and evaluate HyP-NeRF on three tasks: generalization, compression, and retrieval, demonstrating our state-of-the-art results.

Discovering object-centric representations from images has the potential to greatly improve the robustness, sample efficiency and interpretability of machine learning algorithms. Current works on multi-object images typically follow a generative approach that optimizes for input reconstruction and fail to scale to real-world datasets despite significant increases in model capacity. We address this limitation by proposing a novel method that leverages feature connectivity to cluster neighboring pixels likely to belong to the same object. We further design two object-centric regularization terms to refine object representations in the latent space, enabling our approach to scale to complex real-world images. Experimental results on simulated, real-world, complex texture and common object images demonstrate a substantial improvement in the quality of discovered objects compared to state-of-the-art methods, as well as the sample efficiency and generalizability of our approach. We also show that the discovered object-centric representations can accurately predict key object properties in downstream tasks, highlighting the potential of our method to advance the field of multi-object representation learning.

We study multi-task representation learning for the problem of pure exploration in bilinear bandits. In bilinear bandits, an action takes the form of a pair of arms from two different entity types and the reward is a bilinear function of the known feature vectors of the arms. In the \textit{multi-task bilinear bandit problem}, we aim to find optimal actions for multiple tasks that share a common low-dimensional linear representation. The objective is to leverage this characteristic to expedite the process of identifying the best pair of arms for all tasks. We propose the algorithm GOBLIN that uses an experimental design approach to optimize sample allocations for learning the global representation as well as minimize the number of samples needed to identify the optimal pair of arms in individual tasks. To the best of our knowledge, this is the first study to give sample complexity analysis for pure exploration in bilinear bandits with shared representation. Our results demonstrate that by learning the shared representation across tasks, we achieve significantly improved sample complexity compared to the traditional approach of solving tasks independently.

In this paper, we develop an active learning framework for the bipartite ranking problem. Motivated by numerous applications, ranging from supervised anomaly detection to credit-scoring through the design of medical diagnosis support systems, and usually formulated as the problem of optimizing (a scalar summary of) the ROC curve, bipartite ranking has been the subject of much attention in the passive context. Various dedicated algorithms have been recently proposed and studied by the machine-learning community. In contrast, active bipartite ranking rule is poorly documented in the literature. Due to its global nature, a strategy for labeling sequentially data points that are difficult to rank w.r.t. to the others is required. This learning task is much more complex than binary classification, for which many active algorithms have been designed. It is the goal of this article to provide a rigorous formulation of such a selective sampling approach. We propose a dedicated algorithm, referred to as active-rank, which aims to minimise the distance between the ROC curve of the ranking function built and the optimal one, w.r.t. the sup norm. We show that, for a fixed confidence level $\epsilon$ and probability $\delta$, active-rank is PAC$(\epsilon,\delta)$. In addition, we provide a problem dependent upper bound on the expected sampling time of active-rank and also demonstrate a problem dependent lower bound on the expected sampling time of any PAC$(\epsilon,\delta)$ algorithm. Beyond the theoretical analysis carried out, numerical results are presented, providing strong empirical evidence of the performance of the algorithm proposed, which compares favorably with more naive approaches.

Cooperative multi-agent reinforcement learning (MARL) is a challenging task, as agents must learn complex and diverse individual strategies from a shared team reward. However, existing methods struggle to distinguish and exploit important individual experiences, as they lack an effective way to decompose the team reward into individual rewards. To address this challenge, we propose DIFFER, a powerful theoretical framework for decomposing individual rewards to enable fair experience replay in MARL. By enforcing the invariance of network gradients, we establish a partial differential equation whose solution yields the underlying individual reward function. The individual TD-error can then be computed from the solved closed-form individual rewards, indicating the importance of each piece of experience in the learning task and guiding the training process. Our method elegantly achieves an equivalence to the original learning framework when individual experiences are homogeneous, while also adapting to achieve more muscular efficiency and fairness when diversity is observed. Our extensive experiments on popular benchmarks validate the effectiveness of our theory and method, demonstrating significant improvements in learning efficiency and fairness. Code is available in supplement material.

In this paper, we propose an online convex optimization approach with two different levels of adaptivity. On a higher level, our approach is agnostic to the unknown types and curvatures of the online functions, while at a lower level, it can exploit the unknown niceness of the environments and attain problem-dependent guarantees. Specifically, we obtain $\mathcal{O}(\log V_T)$, $\mathcal{O}(d \log V_T)$ and $\hat{\mathcal{O}}(\sqrt{V_T})$ regret bounds for strongly convex, exp-concave and convex loss functions, respectively, where $d$ is the dimension, $V_T$ denotes problem-dependent gradient variations and the $\hat{\mathcal{O}}(\cdot)$-notation omits $\log V_T$ factors. Our result not only safeguards the worst-case guarantees but also directly implies the small-loss bounds in analysis. Moreover, when applied to adversarial/stochastic convex optimization and game theory problems, our result enhances the existing universal guarantees. Our approach is based on a multi-layer online ensemble framework incorporating novel ingredients, including a carefully designed optimism for unifying diverse function types and cascaded corrections for algorithmic stability. Notably, despite its multi-layer structure, our algorithm necessitates only one gradient query per round, making it favorable when the gradient evaluation is time-consuming. This is facilitated by a novel regret decomposition equipped with carefully designed surrogate losses.

Advancements in prompt tuning of vision-language models have underscored their potential in enhancing open-world visual concept comprehension. However, prior works only primarily focus on single-mode (only one prompt for each modality) and holistic level (image or sentence) semantic alignment, which fails to capture the sample diversity, leading to sub-optimal prompt discovery. To address the limitation, we propose a multi-mode token-level tuning framework that leverages the optimal transportation to learn and align a set of prompt tokens across modalities. Specifically, we rely on two essential factors: 1) multi-mode prompts discovery, which guarantees diverse semantic representations, and 2) token-level alignment, which helps explore fine-grained similarity. Consequently, the similarity can be calculated as a hierarchical transportation problem between the modality-specific sets. Extensive experiments on popular image recognition benchmarks show the superior generalization and few-shot abilities of our approach. The qualitative analysis demonstrates that the learned prompt tokens have the ability to capture diverse visual concepts.

Addressing decision-making problems using sequence modeling to predict future trajectories shows promising results in recent years. In this paper, we take a step further to leverage the sequence predictive method in wider areas such as long-term planning, vision-based control, and multi-task decision-making. To this end, we propose a method to utilize a diffusion-based generative sequence model to plan a series of milestones in a latent space and to have an agent to follow the milestones to accomplish a given task. The proposed method can learn control-relevant, low-dimensional latent representations of milestones, which makes it possible to efficiently perform long-term planning and vision-based control. Furthermore, our approach exploits generation flexibility of the diffusion model, which makes it possible to plan diverse trajectories for multi-task decision-making. We demonstrate the proposed method across offline reinforcement learning (RL) benchmarks and an visual manipulation environment. The results show that our approach outperforms offline RL methods in solving long-horizon, sparse-reward tasks and multi-task problems, while also achieving the state-of-the-art performance on the most challenging vision-based manipulation benchmark.

Continual learning requires the model to learn multiple tasks in a sequential order. To perform continual learning, the model must possess the abilities to maintain performance on old tasks (stability) and adapt itself to learn new tasks (plasticity). Task-agnostic problem in continual learning is a challenging problem, in which task identities are not available in the inference stage and hence the model must learn to distinguish all the classes in all the tasks. In task-agnostic problem, the model needs to learn two new objectives for learning a new task, including distinguishing new classes from old classes and distinguishing between different new classes. For task-agnostic problem, replay-based methods are commonly used. These methods update the model with both saved old samples and new samples for continual learning. Most existing replay-based methods mix the two objectives in task-agnostic problem together, inhibiting the models from achieving a good trade-off between stability and plasticity. In this paper, we propose a simple yet effective method, called loss decoupling (LODE), for task-agnostic continual learning. LODE separates the two objectives for the new task by decoupling the loss of the new task. As a result, LODE can assign different weights for different objectives, which provides a way to obtain a better trade-off between stability and plasticity than those methods with coupled loss. Experiments show that LODE can outperform existing state-of-the-art replay-based methods on multiple continual learning datasets.

To solve the spatial problems of mapping, localization and navigation, the mammalian lineage has developed striking spatial representations. One important spatial representation is the Nobel-prize winning grid cells: neurons that represent self-location, a local and aperiodic quantity, with seemingly bizarre non-local and spatially periodic activity patterns of a few discrete periods. Why has the mammalian lineage learnt this peculiar grid representation? Mathematical analysis suggests that this multi-periodic representation has excellent properties as an algebraic code with high capacity and intrinsic error-correction, but to date, synthesis of multi-modular grid cells in deep recurrent neural networks remains absent. In this work, we begin by identifying key insights from four families of approaches to answering the grid cell question: dynamical systems, coding theory, function optimization and supervised deep learning. We then leverage our insights to propose a new approach that elegantly combines the strengths of all four approaches. Our approach is a self-supervised learning (SSL) framework - including data, data augmentations, loss functions and a network architecture - motivated from a normative perspective, with no access to supervised position information. Without making assumptions about internal or readout representations, we show that multiple grid cell modules can emerge in networks trained on our SSL framework and that the networks generalize significantly beyond their training distribution. This work contains insights for neuroscientists interested in the origins of grid cells as well as machine learning researchers interested in novel SSL frameworks.

Regardless of the particular task we want to perform in an environment, there are often shared safety constraints we want our agents to respect. For example, regardless of whether it is making a sandwich or clearing the table, a kitchen robot should not break a plate. Manually specifying such a constraint can be both time-consuming and error-prone. We show how to learn constraints from expert demonstrations of safe task completion by extending inverse reinforcement learning (IRL) techniques to the space of constraints. Intuitively, we learn constraints that forbid highly rewarding behavior that the expert could have taken but chose not to. Unfortunately, the constraint learning problem is rather ill-posed and typically leads to overly conservative constraints that forbid all behavior that the expert did not take. We counter this by leveraging diverse demonstrations that naturally occur in multi-task setting to learn a tighter set of constraints. We validate our method with simulation experiments on high-dimensional continuous control tasks.

In this paper, we take a new look at real-world image super-resolution (real-SR) from a multi-task learning perspective. We demonstrate that the conventional formulation of real-SR can be viewed as solving multiple distinct degradation tasks using a single shared model. This poses a challenge known as task competition or task conflict in multi-task learning, where certain tasks dominate the learning process, resulting in poor performance on other tasks. This problem is exacerbated in the case of real-SR, due to the involvement of numerous degradation tasks. To address the issue of task competition in real-SR, we propose a task grouping approach. Our approach efficiently identifies the degradation tasks where a real-SR model falls short and groups these unsatisfactory tasks into multiple task groups. We then utilize the task groups to fine-tune the real-SR model in a simple way, which effectively mitigates task competition and facilitates knowledge transfer. Extensive experiments demonstrate our method achieves significantly enhanced performance across a wide range of degradation scenarios.

As trained intelligent systems become increasingly pervasive, multiagent learning has emerged as a popular framework for studying complex interactions between autonomous agents. Yet, a formal understanding of how and when learners in heterogeneous environments benefit from sharing their respective experiences is far from complete. In this paper, we seek answers to these questions in the context of linear contextual bandits. We present a novel distributed learning algorithm based on the upper confidence bound (UCB) algorithm, which we refer to as H-LINUCB, wherein agents cooperatively minimize the group regret under the coordination of a central server. In the setting where the level of heterogeneity or dissimilarity across the environments is known to the agents, we show that H-LINUCB is provably optimal in regimes where the tasks are highly similar or highly dissimilar.

Linear scalarization, i.e., combining all loss functions by a weighted sum, has been the default choice in the literature of multi-task learning (MTL) since its inception. In recent years, there is a surge of interest in developing Specialized Multi-Task Optimizers (SMTOs) that treat MTL as a multi-objective optimization problem. However, it remains open whether there is a fundamental advantage of SMTOs over scalarization. In fact, heated debates exist in the community comparing these two types of algorithms, mostly from an empirical perspective. To approach the above question, in this paper, we revisit scalarization from a theoretical perspective. We focus on linear MTL models and study whether scalarization is capable of fully exploring the Pareto front. Our findings reveal that, in contrast to recent works that claimed empirical advantages of scalarization, scalarization is inherently incapable of full exploration, especially for those Pareto optimal solutions that strike the balanced trade-offs between multiple tasks. More concretely, when the model is under-parametrized, we reveal a multi-surface structure of the feasible region and identify necessary and sufficient conditions for full exploration. This leads to the conclusion that scalarization is in general incapable of tracing out the Pareto front. Our theoretical results partially answer the open questions in Xin et al. (2021), and provide a more intuitive explanation on why scalarization fails beyond non-convexity. We additionally perform experiments on a real-world dataset using both scalarization and state-of-the-art SMTOs. The experimental results not only corroborate our theoretical findings, but also unveil the potential of SMTOs in finding balanced solutions, which cannot be achieved by scalarization.